# Day 1 14:00: Panel: Bandwidth Scaling for AI Interconnect – More Wavelengths vs More Fiber?

https://www.youtube.com/watch?v=hqpnL28ka0M

[00:01] and now I will
[00:02] invite Katherine to come in from
[00:07] backstage and uh she will introduce the
[00:11] panelists but I get to introduce
[00:13] Katherine um I'm so pleased that she
[00:16] accepted my invitation to chair this
[00:19] panel uh Katherine got her PhD in the
[00:23] notoriously famous opto Electronics
[00:26] Research opto Electronics Research
[00:27] Center at the University of Southampton
[00:29] is produc
[00:31] stars like Catherine and and others um
[00:33] and then she's really spent much of her
[00:35] career in industry in the last decade or
[00:37] so at Facebook and meta or you know I
[00:41] tended to talk to her to get feedback on
[00:44] on some important questions about where
[00:46] uh you know fiber optic technology would
[00:48] apply and what sort of choices were
[00:51] needed um she uh now has her own um sort
[00:55] of optical uh consulting company arabel
[00:58] systems um and she holds an appointment
[01:02] at another University famous for its
[01:04] Optical research and that's University
[01:06] of California at Santa Barbara UCSB so
[01:10] it is my absolute pleasure to have
[01:12] Katherine join us to moderate the panel
[01:13] and you can see that she has put
[01:14] together a star studded panel so
[01:17] Katherine the floor is yours and I will
[01:19] put on your slides yes well thank you
[01:22] very much Dan and uh welcome everyone to
[01:26] the panel this afternoon this is going
[01:28] to be a lot of fun I was backstage with
[01:32] uh with the panelists and uh we already
[01:36] have the questions flying
[01:38] internally so um the topic this
[01:42] afternoon is bandwidth scaling for AI
[01:45] interconnect more wavelengths versus
[01:47] more
[01:48] fiber and um if you'd go to the next
[01:51] slide please Dan let's get right into
[01:53] this and thank you so much Tony for
[01:56] setting up our panel this afternoon you
[01:58] covered this very nicely
[02:00] this is another very simplified view of
[02:04] the network architectures that are going
[02:07] on in um a hyperscale data center that's
[02:11] running AI compute um workloads and here
[02:16] I've drawn um at the top a front end
[02:19] Network that's typically a clone Network
[02:22] today those are all cabled with optical
[02:25] fiber single mode fiber typically and
[02:29] they connect spine switches fabric
[02:30] switches into top of
[02:32] rack um that's more of an established uh
[02:36] use for use case for fiber optics but
[02:39] what's growing as Tony alluded to is
[02:42] this backend Network and they're shown
[02:45] in blue are some of the there are many
[02:48] networks inside the rack but uh
[02:51] connecting together the gpus or
[02:53] accelerators xpu is a is a growing use
[02:57] case and here's an example of what I'm
[03:00] talking about this is um an example from
[03:04] Nvidia this is their latest uh Blackwell
[03:07] chip and the dgx system that that goes
[03:10] into and in this one rack they have 72
[03:15] gpus and then connecting all of those
[03:18] gpus takes this waterfall of copper
[03:21] cables to connect the back plane there
[03:23] are over 5,000 copper cables and I think
[03:26] you get a sense for um for the volume
[03:30] for maybe the weight of of those
[03:33] connections and uh I think you can see
[03:36] that this might be a a very good
[03:38] application for Optics especially as
[03:41] these uh the number of gpus grows beyond
[03:43] what you can put inside a rack so next
[03:46] slide
[03:51] please I'm making Dan do the hard work
[03:53] thank you so here's some research uh
[03:56] from Del Oro and on the left hand side
[04:00] here the frontend network and the right
[04:02] hand side is backend networks and what
[04:05] Del Oro is um is reporting on here they
[04:09] count the number of Port shipments now
[04:11] this isn't just Optics this is
[04:13] everything this is the port shipments in
[04:15] a network and so in a frontend network
[04:18] you can see this very steady growth in
[04:21] the number of Port shipments that's
[04:23] important that's just the number but
[04:26] then um the the data rate
[04:30] that's is also transitioning and you can
[04:33] see there by 2025 800 gbits per second
[04:36] is the dominant Port speed in backend
[04:40] networks it's interesting these there's
[04:43] some a couple of different Trends first
[04:44] of all the number of ports being shipped
[04:48] is increasing much more
[04:50] rapidly that's important and then the
[04:53] pace at which we're moving through these
[04:56] different speeds is also accelerating
[04:59] and so while maybe in 2025 800 gig is
[05:03] also dominant we move through 800 gig to
[05:07] 1.6t very quickly and so these these
[05:10] applications are
[05:12] diverging next
[05:14] slide and this is an example of some of
[05:18] the performance metrics I'm sure we're
[05:19] going to hear more about this with the
[05:21] panelists discussions today but just to
[05:24] frame things
[05:26] Loosely um interconnect today is 8 100
[05:30] Gig but that's diverging reach the front
[05:33] end is the whole Data Center and the
[05:36] back end is rack scale power efficiency
[05:39] is very important as you heard from
[05:42] Tony's talk 20 P per bit would be the
[05:47] the power efficiency required for an
[05:49] optical module in the front end much
[05:51] lower in the back end and the cost
[05:53] targets are more aggressive in the back
[05:55] end as well this is currently a copper
[05:58] solution so much cheap cheaper
[06:01] expectation um and next slide
[06:05] please so you saw that the transition
[06:09] data rate transition is is accelerating
[06:12] so how do we do this in practice in
[06:14] Optics we have a number of axes we can
[06:17] push
[06:17] on to begin with you push on symbol rate
[06:21] but at 112 gigboard we're really pushing
[06:24] what's
[06:25] possible and already in frontend
[06:28] Networks we are using Pam 4 uh so we're
[06:32] packing more bits per symbol Pam 4
[06:35] allows you to have two bits per symbol
[06:38] and we could push on that some more with
[06:40] more complex modulation schemes but it
[06:44] does come at the expense of power and
[06:47] complexity so let's look at the number
[06:49] of fibers and the number of wavelengths
[06:51] and we have another two axes and already
[06:54] in the front end
[06:55] networks um parallel single mode fiber
[06:58] is an example of solution that has four
[07:02] fibers in each
[07:04] Direction and uh cwdm4 is an example of
[07:09] having four wavelengths in each
[07:11] Direction so this has already been done
[07:13] and the question to the panel is what's
[07:16] next because we got to keep going we got
[07:18] to keep doubling keep doubling and so
[07:21] what's the right parameter here to keep
[07:23] pushing on so we have a fantastic panel
[07:28] um next slide please
[07:30] and um these are World experts in their
[07:35] fields and I'm very happy that they
[07:37] accepted the invitation to come and talk
[07:39] about their particular view of what the
[07:41] right way to solve this problem is I'd
[07:44] like to ask Dan CA to come
[07:49] forward and um join us front of stage hi
[07:55] Dan and um so Dan is at IBM uh at the
[07:59] Watson Research Center he's been there
[08:01] for very many decades and has done some
[08:04] fantastic work in a number of different
[08:06] areas of fonics so please Dan share with
[08:09] us your perspective on this interesting
[08:12] challenge so many
[08:15] decades uh uh next slide please thank
[08:20] you for inviting me to speak um
[08:22] Katherine and Dan so I'm going to uh I'm
[08:25] going to talk mostly about fiber and a
[08:28] little bit of about wavelength I'll give
[08:30] you my thoughts on wavelength and to go
[08:33] and talk about some past examples of
[08:36] various projects that um tried to use a
[08:38] lot more fiber um and I'm talking about
[08:41] single core ribbon fiber and multicore
[08:44] multi-core ribbons um but I'll show you
[08:47] that the market has not been kind to
[08:51] um high fiber count in fact it appears
[08:54] to favor favor fewer fibers and higher
[08:57] speeds um next slide
[09:02] so uh the past is sort of speckled with
[09:06] lots of high fiber count
[09:09] transceivers um and mostly multimode
[09:11] transceivers so these would be the type
[09:15] that you would want to use in a in a
[09:16] backend network uh we had xan Optics uh
[09:20] producing a 72 fiber module uh
[09:24] airite 36 fiber teral link
[09:27] 48 um none of those really got Market
[09:31] traction the one that did get Market
[09:33] traction was the snap
[09:35] 12 um and maybe because that's it had an
[09:39] MSA but that came out of H the Packard
[09:42] following their Polo and pony projects
[09:45] and it had only 12 fibers so in general
[09:49] high fiber count Optical links that's
[09:52] fine you can leave it on that next page
[09:54] uh go ahead next
[09:56] page uh this is an example more recent
[09:59] example this one had 336 fibers and I I
[10:03] don't believe they're in business
[10:05] anymore um but in general the market has
[10:08] um not been good or not been kind to
[10:11] high fiber count uh
[10:14] transceivers uh next page
[10:17] please so even within IBM we pursued uh
[10:21] several projects with high Channel count
[10:24] as a as a way to crank up the data rate
[10:27] so here's our terbus example
[10:30] um terbus had 48 channels so 24 transmit
[10:34] 24 receive um what made terbus unique
[10:37] was we coupled two polymer Optical wave
[10:39] guides on a printed circuit board uh the
[10:42] one thing I'll say about polymer Optical
[10:44] wave guides is you can make them cheap
[10:47] you can fabricate lots of them
[10:49] lithographically but adding them to a
[10:52] printed circuit board invites a whole
[10:53] host of other issues um that that make
[10:57] them difficult to implement and next
[10:59] slide please so after polymer wave
[11:02] guides we went back to a high fiber
[11:05] count so we just did a 48
[11:07] Channel um device with a 4x12 uh fiber
[11:13] array uh with a tight Bend
[11:15] radius um next slide and then trying to
[11:19] get away from high fiber count we went
[11:22] to multi-core devices so essentially we
[11:25] took the same 48 Channel device uh but
[11:28] we rep placed it with
[11:32] um six uh or um I'm sorry eight six
[11:37] Channel multicore fiber and so you know
[11:40] we looked into could you have
[11:43] arrays uh like MTP connectors with uh
[11:46] populated with multicore fiber U we
[11:49] studied that for a while and and um it's
[11:52] possible to build those uh it's possible
[11:55] to put all your opto electronics uh you
[11:59] know sort of match a multicore fiber
[12:01] array in a butt coupled manner um it
[12:04] becomes really tricky to wire up the
[12:06] circuitry um to the same level um and
[12:11] and then in the end we didn't see a cost
[12:14] Advantage multicore fiber invites its
[12:16] own
[12:17] issues uh next slide
[12:21] please uh oh here's another example we
[12:24] also it wasn't just six core hexagonal
[12:26] we looked at um looked at four core
[12:29] Square as well uh next slide
[12:33] please um and I want to say something
[12:35] about yield um and uh in the motion
[12:39] project we had done a 16 channel
[12:42] transceiver but if you look at the
[12:44] picture of uh all the D on the glass
[12:47] substrate you'll notice uh certainly in
[12:50] the photo dial that there's sort of four
[12:52] different shades of of gray there and
[12:55] and what that really is is we found that
[13:00] um while in principle you can make
[13:02] arrays of 16 devices uh what we found in
[13:06] practice was that four arrays of four
[13:09] channels uh yielded better than than one
[13:12] large array and that's including
[13:14] assembly and test and that's sort of
[13:17] where I'm thinking that some of this
[13:19] needs to go
[13:21] is um you know
[13:23] numerous
[13:25] small transceivers with with lower
[13:28] um when we go to High Channel count um
[13:29] uh next slide
[13:31] please
[13:38] we need to consider
[13:40] reliability um
[13:42] and for the motion project uh which had
[13:46] 16 channels um our approach to
[13:49] reliability was to use two to one
[13:51] sparing and the way we did so first
[13:54] thing just
[13:55] theoretically two to one sparing will
[13:58] buy you about three orders of magnitude
[14:00] Improvement in
[14:03] reliability um it's it's pretty powerful
[14:06] uh the way we implemented it was our 1x4
[14:09] array of vixel was really a
[14:11] 2x4 where alternating rows had
[14:15] orthogonal
[14:16] polarization um and then I'm just
[14:19] showing a little sort of Ray tracing
[14:21] diagram how we coupled the two
[14:24] polarizations uh into one fiber um so it
[14:29] doesn't really matter what you do or
[14:31] what technology you have um sparing I
[14:35] think especially for high Channel count
[14:37] I think sparing uh is something that has
[14:40] to be um
[14:42] included next chart please and then for
[14:46] high fiber count uh connectors um so in
[14:51] the market today you can get a 72 fiber
[14:54] MTP from us
[14:56] connect um for aava
[14:59] uh at one point almost 10 years ago made
[15:03] a 96 fiber uh
[15:06] connector that's not commercially
[15:08] available um both of these fibers are
[15:11] multi mode both of these connectors are
[15:13] multi mode though um what I'm seeing in
[15:16] single mode what's available today is up
[15:19] to 32 fiber two rows to 16 I'm told that
[15:23] a 48 fiber is on the horizon so two rows
[15:28] of 24 where they dro the pitch to
[15:32] 165 um but over on the right is is kind
[15:36] of a complicated chart but the message
[15:38] of the chart is that the more fibers you
[15:42] put in a connector um the higher the
[15:44] cost is in fact a 72
[15:48] fiber uh MTP 72 cable is is about eight
[15:52] times the cost of six 12 fiber cables so
[15:56] same same fiber count it's just it's
[15:59] just a lot more expensive and the same
[16:01] holds true for the single mode case an
[16:04] MTP 32 is about three times the cost of
[16:08] two mtp1
[16:10] 16s um and part of this is that the
[16:13] piece Parts themselves are more
[16:16] expensive um it's more labor intensive
[16:19] to assemble these high fiber count
[16:21] connectors and the yield is lower and
[16:25] and the other thing that I find going on
[16:27] in our industry
[16:29] we have a lot of work to automate
[16:32] transceiver assembly and test but our
[16:35] fiber connectors and cables still
[16:37] involve a substantial amount of manual
[16:40] labor and so I think that's sort
[16:44] of goes against moving to high high
[16:48] count fiber
[16:50] connectors uh next page
[16:53] please so a couple of comments on
[16:56] wavelengths I feel that a small number
[16:59] of additional wavelengths like two or
[17:01] four is relatively easy to add um
[17:05] without much additional cost course
[17:08] wavelength spacing avoids temperature
[17:11] control um a large wavelength count
[17:14] implies a small wavelength spacing and
[17:17] so that then requires some degree of
[17:19] temperature control or wavelength
[17:21] control uh which is going to cost some
[17:25] amount of
[17:26] energy and I've tried to make the point
[17:29] that reliability drives some sparing
[17:33] requirements um I don't have a chart to
[17:35] show this but but wave links are harder
[17:37] to spare than than fibers or or other
[17:41] things that share a fiber um I'm sort of
[17:45] putting out a question but how do you
[17:47] spare a comb laser because when a when a
[17:50] comb laser goes down it's not like it
[17:52] takes out one wavelength it takes out
[17:55] all of its wavelengths and and that
[17:57] seems to be a single point of failure
[17:59] uh for many channels so um that's a
[18:04] that's a concern for for adding a lot
[18:06] more
[18:07] wavelengths next slide
[18:10] please so just some additional comments
[18:15] um you know there are some advantages to
[18:17] adding more fiber um you know you can
[18:20] have lower cross talk and that would be
[18:21] lower cross talk compared to say
[18:24] wdm uh I don't have a slide for this but
[18:27] IBM spent a lot of time looking at
[18:30] parallel buses with a forwarded clock
[18:33] meaning that we're devoting uh an extra
[18:36] fiber to sending a forwarded clock um it
[18:40] does lead to some tighter skew and
[18:41] Jitter requirements but it leads to
[18:44] lower energy and lower latency and lower
[18:47] beer um I did mention the polymer wave
[18:50] guides can be made cheap so you can add
[18:53] a lot of them but they they invite a
[18:55] whole host of of other issues and then I
[18:59] the disadvantages of more Fiers is the
[19:02] fiber bulk um maybe multi-core fiber can
[19:05] offer some release relief there um but I
[19:09] try to drive home the point that there's
[19:11] a big yield disadvantage it's really
[19:14] hard to yield a large number of optical
[19:17] devices and it's harder to yield High
[19:19] count fiber connectors the reliability
[19:22] does scale to the power Bend and you
[19:25] definitely need a sparing and redundancy
[19:28] scheme
[19:29] okay I think this is the last slide next
[19:31] one no oh one more slide okay Katherine
[19:35] wanted us to talk about energy my my
[19:37] quick remark about energy is Optical
[19:41] links are essentially analog in nature
[19:44] and so when we look at Analog
[19:48] seos um they have this this property
[19:51] that I'm showing in the upper right uh
[19:54] which is there's a certain minimum
[19:55] energy just to turn on the circuits and
[19:58] below that minimum energy the energy per
[20:01] bit uh goes up and then above that
[20:04] minimum energy the energy per bit also
[20:06] goes up because you're hitting the so on
[20:09] the left side it's going up because you
[20:11] have fewer bits that you're advertising
[20:13] the energy over and on the right side
[20:15] it's going up is you have to push more
[20:18] current in to overcome the bandwidth
[20:19] limits so each technology node has a
[20:25] different Optimum data rate that
[20:27] minimizes energy consumption there's no
[20:30] like one correct answer it really
[20:32] depends on the technology node that you
[20:35] choose and also going slower generally
[20:38] does not improve the energy per bit once
[20:40] you're below the data rate of the
[20:43] minimum okay next slide this is the last
[20:46] one
[20:48] so in in summary um I didn't have a side
[20:53] that says this but Optics cost is what
[20:56] really dominates everything dominates
[20:58] all the decisions cost appears to be
[21:01] minimized when you are running your
[21:04] Optics at the highest data rate that
[21:05] they can
[21:06] support um I don't have a strong opinion
[21:10] on more fibers versus more wavelengths
[21:12] because adding either adds cost although
[21:16] I I really think until they start
[21:19] automating fi fiber connector
[21:22] assemblies and and improving on the
[21:24] yield adding more fibers I think is is
[21:26] more challenging than adding more
[21:27] wavelengths
[21:29] I do think that this suggests a large
[21:32] collection of medium Lane count
[21:35] transceivers uh might be Optimum
[21:38] especially if they have very small
[21:39] packages and can be
[21:42] um packaged right next to each other and
[21:45] then of course the real
[21:47] Optimum uh is going to depend on the
[21:50] exact application and the technology
[21:52] node that you
[21:54] choose okay
[21:58] well thank you very much Dan that was uh
[22:01] very interesting and I'm looking to see
[22:05] I actually don't see any questions and
[22:08] since you did share um quite a lot of
[22:12] information which I think is wonderful
[22:13] thank you for uh talking about cost and
[22:16] energy and sharing your experience over
[22:19] over many different types of
[22:21] Technologies that's great so um I saw
[22:24] one question from
[22:27] Jeff but I don't I don't know how to
[22:30] answer it I don't know the answer to
[22:32] your question
[22:35] Jeff cther do these questions come
[22:38] through see Jeff's question are you
[22:41] seeing it on the slack Channel or are
[22:43] you seeing it in the chat which is I'm
[22:45] seeing it in the question and answer on
[22:48] Zoom really anyway so Jeff Hutchinson's
[22:51] question is do you see a road map to
[22:53] automating fiber connectors in the
[22:56] industry and that's a good question I
[23:00] I'm not the right person to answer that
[23:02] question um but I feel that the lack of
[23:06] or I feel that the amount of manual
[23:08] labor in cabling and
[23:11] connectorization um is is is something
[23:14] that really needs
[23:17] addressing as much as automating
[23:20] transceiver assembly
[23:21] does there's another question before
[23:23] that how do the types of limitations map
[23:25] to New Optical transceiver Technologies
[23:27] such as m
[23:29] LEDs I think we'll have that'll be
[23:31] addressed by some future panel I'll let
[23:34] B answer that and I I do see one other
[23:38] questions just come through the the
[23:39] slack Channel and since you mentioned it
[23:41] a few times let's clarify what do you
[23:43] mean by medium count
[23:46] transceiver so so medium count um if I
[23:51] look
[23:52] at
[23:54] where
[23:56] um single mode
[23:59] seems to be they're getting the 16
[24:02] channel connector done about right
[24:05] similarly
[24:06] for um multi mode although
[24:09] weirdly the the 1 by 16 ends up being a
[24:13] little bit more costly than the
[24:16] 2x4 I'm not sure why that is um so I
[24:21] think in in that range maybe the sub 32
[24:25] Channel range 16 to 32 is where I would
[24:29] think of as being
[24:32] medium
[24:34] great thank you very much
[24:37] Dan um we'll have Dan um come back for
[24:41] the panel Q&A later on and I'd like to
[24:43] call Peter Peter Windsor would you join
[24:47] us front of stage so um Peter has had a
[24:50] a long career at Bell laabs and has um
[24:55] focused on uh SDM space division
[24:58] multiplexing more recently of course in
[25:01] his work with submarine you know all
[25:03] about multiple wavelengths um and here
[25:07] he is with nubis focusing on Datacom so
[25:11] let's tell us about how you see the
[25:13] right solution
[25:14] Peter thanks a lot Katherine thanks for
[25:17] your introduction and for inviting me to
[25:19] this panel so I'll talk about high
[25:21] speeds over lots of parallel fiber and I
[25:24] think that's the best for optical data
[25:26] com IO um as us you shall see so next
[25:29] slide then please so first of all we all
[25:32] need to remember that uh every
[25:34] communication is high speeed because any
[25:37] processor iio operates at a multiplex of
[25:40] the clock frequency the clock
[25:41] frequencies are one to two gigahertz for
[25:43] processes and um everything that we do
[25:46] in electrical and Optical operates at
[25:48] many gigabits per second what you see
[25:50] here in this table is a summary of three
[25:53] prevalent um electrical interfaces hbm3
[25:57] ucie and serd
[25:58] um that go from all the way 6 gbits per
[26:01] second to 200 gbits per second for for
[26:04] mod than 30s and uh which speed you
[26:08] choose depends on a couple of different
[26:10] trade-offs uh some of them are related
[26:12] to the processor itself in terms of how
[26:14] much area in the chip do you want to
[26:16] spend for your IO macros and also the
[26:19] breakout of the processor what's your
[26:21] Bump density what's your packaging
[26:22] strategy and your packaging complexity
[26:25] so that's one part that determines the
[26:27] Sol solution you're going to choose and
[26:30] another one is the transmission medium
[26:31] itself like what Lane densities are we
[26:34] talking about are we talking substrates
[26:36] pcbs maybe even twin ax cables what
[26:39] distance do I want to achieve and what
[26:41] complexity am I willing to afford on my
[26:43] transmission medium so all of these
[26:46] represent a trade-off space that we need
[26:48] to consider when we talk about um
[26:50] interface Technologies and speed but one
[26:53] thing is for sure if you want to go
[26:55] beyond the centimeter case uh that is
[26:58] best supported by high-speed Surs that
[27:00] has been true and is true today um next
[27:04] slide please and a good thing is that
[27:07] Optics likes fast so Optical channels
[27:10] support high speeds because they don't
[27:12] have this rolloff that you see on
[27:14] electrical channels Optics is by nature
[27:17] broadband and the good news is and Tony
[27:19] alluded to that in his keynote as well
[27:21] that modern series can directly Drive
[27:24] Optics especially if those Optics are
[27:27] built to look l to look like an
[27:29] electrical Trace to the CIS then the CES
[27:31] can directly Drive the Optics you don't
[27:33] need any signal reconditioning or R
[27:36] timing um um in your Optical module
[27:39] whatever that module might be um you can
[27:42] also use converted chiplets uh that's
[27:44] again as Tony said in in his keynote uh
[27:47] like a UCI 230s converter if you want to
[27:49] take advantage of the chiplet ecosystem
[27:52] and you want to have your processor
[27:54] speak ucie to a converter chiplet that
[27:57] then converts that to high speeded and
[27:59] then same story applies you go highp
[28:01] speed over Optics and these two
[28:03] solutions by the way consume the same
[28:05] power uh end to end it's a wash next
[28:08] slide
[28:09] please okay now let's talk about slow
[28:12] Optics versus Fast Optics a little bit
[28:14] so um the advantage of fast Optics is
[28:18] that you can directly drive it with
[28:19] 30ies the advant the disadvantage of
[28:21] slow Optics is that slow processor iOS
[28:25] cannot directly drive slow Optics
[28:27] typically typically what you do is you
[28:30] use some sort of digital signal
[28:32] conditioning retiming or G boxing as
[28:34] part of your Optical engine that needs
[28:36] to reestablish uh the signal quality
[28:38] because these die to die interfaces are
[28:40] not meant to drive Optics uh at at the
[28:43] low speeds also slow Optics need much
[28:47] more components you need more modulators
[28:50] more photo detectors which becomes the
[28:52] size and the yield issue um uh as as Dan
[28:56] also said before and also o slow Optics
[28:59] fully amortized perlane overheads
[29:01] example for that is modulator bias for
[29:03] example which is a fixed cost
[29:05] essentially that you have for a channel
[29:06] if it's a slow Channel or a fast Channel
[29:08] it doesn't matter you need that bias and
[29:11] another one is fiber density you can't
[29:13] afford to send 25 gigabits uh over
[29:16] individual fibers because the density
[29:18] wouldn't allow it so you need
[29:20] multiplexing like wdm if you speak about
[29:23] uh slow Optics next one
[29:25] please okay so here is a an example just
[29:28] to show and to to dismiss the myth a
[29:31] little bit that slow is better energy
[29:34] efficient than fast um these are
[29:37] actually Hardware demonstrated things
[29:39] this are this is not PowerPoint
[29:40] engineering this comes from people
[29:43] reporting actual in to end experiments
[29:45] at 16 gigabit with eight eight
[29:47] wavelengths and a 100 gigabit with one
[29:49] wavelength both silicon photonics up to
[29:52] couple kilometers reach and both consume
[29:55] about 10 PJ per bit end to end if you
[29:57] count everything in including the 30s in
[29:59] the host processor interestingly the
[30:02] slow solution that was again actually
[30:05] demonstrated shows about 300 gigabits
[30:08] per second per millimeter um whereas the
[30:11] fast um option is in by 500 gigabits per
[30:15] second per millimeter and what do I mean
[30:17] by n by 500 that's shown on the next
[30:20] Slide the the fast option if you
[30:24] Leverage 30ies What it allows you to do
[30:26] is you reach it further back away from
[30:29] the chip you can reach multiple rows of
[30:32] optical engines because your series can
[30:35] drive it there is no problem bridging
[30:37] those longer distances with your seris
[30:39] so if one row has 500 gigabits two rows
[30:43] has one terabit per millimeter four rows
[30:45] has two terabit per uh per millimeter so
[30:48] you can stack up if you do things
[30:50] correctly and if you leverage this cies
[30:53] and the 2D tiling architecture which
[30:56] also means you need to um Escape in the
[30:59] third dimension just as Dan showed with
[31:02] the multim mode uh so escaping in the in
[31:05] the third dimension with a 2d array is
[31:07] definitely a good thing next one
[31:10] please okay so now now we are all
[31:13] convinced that high speed is the right
[31:14] way to go so now the question is should
[31:17] we send high speeds over dedicated
[31:19] Optical fibers like continue the
[31:22] electrical Trace all the way through the
[31:24] optical fiber without uh any change or
[31:26] should we use multiplexing by either wdm
[31:29] polarization multiplexing coherent or
[31:32] even bidirectional transmission so to
[31:35] answer that we all need to acknowledge
[31:37] that multiplexing is a means to amortize
[31:39] expensive channels that's true in all of
[31:41] engineering you only Multiplex if the
[31:44] channel is so expensive that you need to
[31:46] share it among different participants of
[31:48] the communication Paradigm and if you go
[31:50] to the next slide then we'll ask how
[31:52] expensive is a fiber so this this
[31:55] channel summarizes this table summarize
[31:58] is the cost of a fiber all across
[32:01] Optical applications all the way from
[32:03] submarine transpacific cables to mlai
[32:06] clusters and the cost here is given in
[32:09] multiples of a 200g transceiver cost so
[32:14] a transpacific fiber one extra strand of
[32:17] fiber costs 600 times the cost of a 200g
[32:20] transceiver and that of course means in
[32:23] in submarine you need to Multiplex as
[32:25] much as you can to amortise the cost of
[32:27] that
[32:28] so you use coherent quam pmax wdm you
[32:32] throw all multiplexing possibilities at
[32:34] that problem to to amortize that cost
[32:37] but for short distances 10 to 100 meters
[32:41] the cost of an additional fiber is in
[32:42] the single- digit percents of a 200g
[32:46] transceiver so there is no need for
[32:48] multiplexing from a uh from a cost point
[32:51] of view uh so you use parallel fibers
[32:54] that's a good thing to do okay next one
[32:58] so here is a an example that shows how
[33:00] multiplexing also costs you power so
[33:03] what I compare here is an 8x2 200g dr8
[33:07] so one wavelength over eight uh fiber
[33:10] pairs versus an 8X 200g f8 so eight
[33:13] wavelengths over one fiber pair and the
[33:16] e2o and O2 has the same um power
[33:19] consumption because you use the same
[33:20] amount of modulators the same modulators
[33:22] really but the lasers are pretty
[33:24] different because for WM you need to go
[33:26] through a Max dmax pair
[33:28] which if a Max and dmax has one and a
[33:31] half DB loss each that's 3db so you
[33:33] double your laser power and also the
[33:35] efficiency of the lasers tend to be much
[33:37] better if it's just a single wavelength
[33:39] unstabilized laser it's much better uh I
[33:42] here only assume 20% to 15% uh it might
[33:46] be the difference might be even larger
[33:47] than that so that just shows you that
[33:49] multiplexing comes at the cost and
[33:52] that's what we need to acknowledge okay
[33:54] next one please so then there is another
[33:57] aspect to the issue and that's uh
[33:59] networking so usually you want to do
[34:01] full Ric switching you want to break out
[34:04] all your links to multiple end nodes or
[34:06] switches uh to build those uh clone
[34:08] networks that Katherine talked about in
[34:10] the introduction and in order to do that
[34:12] you need a network capable protocol so
[34:14] you need to uh to to use like ethernet
[34:17] infin band Envy link something like that
[34:19] where you can actually do networking you
[34:21] can't just use any proprietary interface
[34:24] or die to die interface that's not meant
[34:26] for networking so that's very important
[34:29] and that's favor single wavelengths and
[34:30] next one uh please so if you Multiplex
[34:33] those up into wdm then you have these
[34:35] big fat pipes and you can't do
[34:37] networking with those big fat pipes then
[34:39] it's an endtoend big fat connection
[34:42] that's not suited for um for any uh
[34:45] networking application next one please
[34:48] so this is just the conclusion uh what
[34:50] NIS uh how NIS solves this problem with
[34:53] the xt600 alra high density Optical IO
[34:56] you see on the left side
[34:58] our near package Optics module with 16x
[35:01] 100 Gig full duplex has a 15x 15 mm
[35:04] footprint and if you gang that together
[35:06] in a staggered array because we can uh
[35:08] in terms of the 2D um arraying we get 10
[35:12] terabit full duplex on the footprint of
[35:14] half a business card uh that's sort of
[35:17] um what we can do today you see in the
[35:19] table all the parameters that matter
[35:21] High IO density low latency low power uh
[35:24] full intack to interr connectivity from
[35:27] one meter to hundreds of meters full
[35:29] Radix networking because we use single
[35:31] wavelength multiple fibers and full
[35:34] Network compatibility with all the uh
[35:36] networking protocols that are out there
[35:42] thanks thank you very much Peter and I'm
[35:46] looking to see if there are any
[35:47] questions in any of the different forums
[35:49] I don't see any oh Dan do you have a
[35:52] question did you see one yeah just no I
[35:54] didn't see one I just H have one very
[35:57] quickly Peter here
[35:58] what is driving these fibers is it
[36:00] lasers or
[36:02] LEDs oh this is laser based this is
[36:04] silicon photonics uh you have lasers
[36:06] that are modulated by silicon phonics
[36:08] modul okay than
[36:13] care and uh I have some questions
[36:17] regarding the the size could you you did
[36:21] mention it but it was kind of fast what
[36:23] the pcie card that you that you show
[36:26] there what is the chip on the right hand
[36:28] side that you're showing up against and
[36:30] what's the scale that we're talking here
[36:32] oh so so this is a this is just a mockup
[36:35] this PCI card is a mockup um it's uh
[36:38] it's a regular size uh PCI card the the
[36:41] modules themselves I I don't know
[36:43] exactly how how big the chip is or the
[36:45] modules but one of our modules is 15 by
[36:48] 15 millimeters millimeters
[36:52] yeah thank you again I don't see any
[36:55] more questions and we are a little
[36:57] behind so let's move on please thank you
[37:00] very much
[37:01] Peter and I would like to call Rebecca
[37:04] to come for yep that was fast thank you
[37:06] Rebecca you're welcome Rebecca shavitz
[37:10] is at mix Tex technology she has a long
[37:14] career you may have seen her um she's
[37:16] certainly got a lot of experience in the
[37:18] industry working at coring and at
[37:20] broadcom so she knows what she's talking
[37:22] about and please share with us what you
[37:25] think the right solution is Rebecca
[37:28] happy to thank you Katherine for the
[37:30] warm introduction um as Katherine said I
[37:34] am at mixed Technologies I'm the
[37:35] co-founder and chief product officer
[37:38] where we are scaling AI
[37:41] sustainably um I also love that the
[37:43] keynote and you know talks before we we
[37:46] talked a lot about how to scale but also
[37:47] how to do it
[37:49] sustainably um in a way that's power
[37:51] effici next slide so before we get into
[37:55] answering the question at hand I do want
[37:57] to restate the problem that we're trying
[38:00] to solve um that's always first and
[38:03] foremost the important aspect of what
[38:05] we're doing here um there's a lot that
[38:08] we've done in history in Optics field
[38:11] but as the market shifts towards these
[38:14] AI models we also have to shift a
[38:18] mindset so when we look at AI
[38:21] specifically some of the things that are
[38:24] called out that Katherine spoke about
[38:26] hot speaker spoke about is these very
[38:28] very large models many many gpus or xus
[38:31] generically that are connected together
[38:34] and that requires inherently a lot of
[38:36] interconnects and if we use today's
[38:39] technology that means we have a lot of
[38:41] overhead on networking because we want
[38:44] these to be full bandwidth non-blocking
[38:47] scenarios and that means a lot more
[38:50] switches a lot more interconnects in
[38:52] order to adjust to
[38:54] that in addition when we add on these
[38:57] networking layers we also add on things
[39:00] like latency and variation to latency
[39:03] and those are compounding our needs for
[39:05] memory access because we need to store
[39:07] more data while we wait for all the data
[39:11] to get through the network itself and
[39:14] that memory also needs interconnects so
[39:17] now we've compounded this problem of
[39:19] interconnects not just in meeting the
[39:21] GPU requirements of talking to each
[39:23] other but also in accessing memory and
[39:26] all of this overhead means that we are
[39:28] exploding our cost basis right for each
[39:31] GPU we have a lot more components that
[39:34] need to work reliably which means we buy
[39:37] even more to have redundancy and we're
[39:39] exploding these costs and we have a huge
[39:42] imbalance of resources going on and
[39:45] decreasing Cloud margins right as we do
[39:48] these you know inquiries into the models
[39:51] and that all Burns a ton of power so how
[39:54] do we solve this problem right we do
[39:55] need to fundamentally solve the massive
[39:57] load of interconnects and we need to do
[40:00] this in a hardware solution next
[40:04] slide not only that we also have this
[40:07] huge diversity in what's needed right
[40:10] meta presented last year at ocp a really
[40:13] great diagram of showing how you know
[40:17] whether you're doing training or you're
[40:19] doing inference you might need very
[40:22] different solutions and then not only
[40:25] that you might be doing a large language
[40:26] model you might be doing recommendation
[40:28] ranking and there might be future things
[40:30] like video processing and things that we
[40:33] haven't yet discovered today as being
[40:35] necessary and these are varying things
[40:38] like model size or scale your memory
[40:40] needs your network
[40:42] latency and one of the things I want to
[40:44] point out is on this plot every single
[40:47] one of those parameters are influenced
[40:50] by your interconnect capabilities other
[40:53] than compute
[40:55] itself so if you need more memory you
[40:58] need to interconnect to more memory
[41:00] right and network is I think very clear
[41:04] um but if you need more models we need
[41:05] model sizes and more nodes we need to
[41:08] interconnect those so these are all
[41:09] incredibly important aspects of how we
[41:12] scale
[41:13] AI next
[41:15] slide so as you know the speakers have
[41:19] mentioned Optics is needed right you're
[41:21] getting to longer distances beyond what
[41:23] a copper can reach so we know Optics is
[41:25] absolutely needed and it's GR
[41:27] substantially because of AI next
[41:30] slide uh I think Marvel had a brilliant
[41:34] graph just showing the level of
[41:37] imbalance in the network when it comes
[41:39] to Optical connectivity right if you
[41:41] take 128 xus as a cluster we're at a one
[41:44] to one overhead when we get to 1 million
[41:47] we're at a 10 to one overhead it's a
[41:50] huge discrepancy in terms of what we
[41:53] need to design toward and the volume
[41:56] that we need to design toward which is
[41:58] very different than history has dictated
[42:00] Optics to
[42:02] drive next
[42:04] slide so we we kind of saw this coming
[42:08] we said oh we're going to need some
[42:10] solution here and we have tried to
[42:13] migrate from Standard Optical plug bles
[42:16] into things like onboard Optics or a
[42:19] substrate based CPO integration and I
[42:22] call that substrate base it's a kind of
[42:24] a 2d integration where we are limited by
[42:27] the substrate density in interconnects
[42:31] versus the PCB density that we're
[42:33] limited by in an optical plugable
[42:36] case these solutions that we've been
[42:38] working toward have provided things like
[42:41] better bandwidth better density better
[42:44] rics but they have not yet been adopted
[42:48] into volume and the question is if AI
[42:51] needs all of this why hasn't it yet been
[42:54] adopted next slide
[42:58] and the answer and and I love that in
[43:00] the keynote this was alluded to
[43:02] packaging is a really big challenge
[43:05] right Dan even mentioned this packaging
[43:08] is hard and it historically has been
[43:10] hard and the more you add fibers into
[43:12] the mix the harder it is and part of it
[43:16] has been our mentality of how do we
[43:19] scale up a transceiver rather than how
[43:22] do we integrate with ASAC technology as
[43:25] it's manufactured
[43:27] today and that's where we need to have a
[43:30] paradigm shift in how we actually
[43:32] approach Optics in this new world next
[43:36] slide and so that's what I call the
[43:38] siliconization of Optics where we are
[43:40] designing Optics in a way that's
[43:44] directly compatible with existing
[43:47] Advanced manufacturing processes without
[43:50] disruption where we have automated
[43:52] processes that don't doesn't need a
[43:54] technician to come in and intercept or
[43:57] active alignment in order to create that
[44:00] solution where we're focused on how
[44:03] testing happens throughout that process
[44:06] how manufacturing happens throughout
[44:08] that process and center the design on
[44:10] those requirements where we stay pitch
[44:13] matched to the fundamental die itself
[44:15] that we want to get that interconnect
[44:16] off so we can scale with the Asic die
[44:20] and we can do all of this at waer scale
[44:22] and that is what's critical in Optics
[44:24] field and creating those Solutions that
[44:27] enable the whole industry to to shift
[44:30] into the volumes that are needed for AI
[44:34] next
[44:35] slide so you're gonna have to step
[44:37] through this with me Dan um if we talk
[44:40] about copper uh
[44:42] next that you know is kind of the
[44:45] fundamental basis right has the lowest
[44:47] radex lowest reach the number of xus you
[44:50] can get in a cluster for a given latency
[44:53] is relatively low it's something like
[44:55] eight right that you can fit on a
[44:57] board next slide now if we add Optical
[45:01] transceivers into that we can increase
[45:04] that reach which is great but we're not
[45:07] increasing the density because we're
[45:09] still limited by the PCB and the io
[45:11] through cages right so we haven't really
[45:13] solved the problem next
[45:16] slide so we can now go to 2D type
[45:20] integration on a standard CPO solution
[45:24] and we can improve radex because we can
[45:27] now have the density of a substrate in
[45:30] terms of the interconnect capacity to
[45:33] your Optical engine so you get some
[45:35] improvement in there and then you can
[45:36] get more xus in a cluster for a given
[45:38] latency that's great what we're driving
[45:41] at at mix
[45:43] next is to really dramatically change
[45:46] that and here's the answer to your
[45:49] question Katherine um I believe that
[45:52] scaling radic is the fundamental that
[45:54] you need to really drive solutions for
[45:57] AI and what they need so with an
[46:01] enhanced radic an Innovative design to
[46:04] make sure that you stay within the
[46:06] advanced manufacturing pipeline you can
[46:08] get something like 800 xus in a cluster
[46:12] that's less than 50 nanc leny that's the
[46:14] equivalent of what we do today on a
[46:16] board with eight right so that's a 100x
[46:19] Improvement we can get things like 500x
[46:22] nearly 500x Improvement in these large
[46:25] model for training
[46:27] which gets us to that 1 million or
[46:29] further Mark or provides redundancy
[46:32] things that have you know challenged the
[46:35] reliability of those
[46:37] networks and with that you're also
[46:39] creating a solution that can integrate
[46:41] whether it's in the back end or the
[46:42] front end of the network because you're
[46:43] a Dr compliant and you can match any of
[46:46] the lower density systems that are
[46:48] already out there that maybe you just
[46:49] want to throw a transceiver into a
[46:51] single box sort of solutions and so in
[46:54] that way you're solving the problem that
[46:56] is actually important for the AI
[46:58] industry yes it makes it challenging yes
[47:01] you have to solve the packaging but
[47:03] that's what we need to drive to as an
[47:05] industry to make this real thank
[47:09] you that's very interesting I have one
[47:12] clarifying question for you when you say
[47:15] increasing
[47:17] radex um are you excluding putting more
[47:20] wavelengths to increase bandwidth but
[47:22] then you you have the opportunity to to
[47:26] have different fibers you know take you
[47:28] to a different endpoint so meet that
[47:30] radic or yeah so radex very specifically
[47:34] to me is based on what you're carrying
[47:37] over the fiber right and I am not
[47:41] against in the future scaling with let's
[47:44] say FR technology where we could add
[47:47] more wavelengths to increase the
[47:48] bandwidth that you have on a given fiber
[47:51] but you have to first solve the
[47:52] fundamental and make sure you have a
[47:54] design space you can work in that has
[47:57] sufficient radex for what's needed in
[47:59] the
[48:00] industry sounds good thank you very much
[48:03] Rebecca and in the interest of time I
[48:06] would like to call our next speaker Amit
[48:09] would you step
[48:10] forward and um Amit nagara is at
[48:15] Intel and recently joined he has a long
[48:18] history in photonics from source
[48:19] photonics and Rocky photonics and he's
[48:23] currently the VP and general manager of
[48:26] Intel silicon fonics group please go
[48:29] ahead thank you Emy tell us about your
[48:36] solution if you could unmute for us
[48:38] please Amit thank
[48:42] you thanks for the opportunity Katherine
[48:44] and Dan if you go to my first content
[48:47] slide it's flip through all the
[48:50] way and if it has some kind of
[48:54] automation there we go yeah so um next
[49:03] please I think uh through this
[49:05] presentation keeping the key question at
[49:07] hand which is you know how do we scale
[49:10] for AI networks is wavelength more
[49:12] important is um fiber count more
[49:15] important I think this is a mix s of
[49:18] synergy of our experience at Intel along
[49:20] with many of the panelists before um
[49:23] having shipped silicon photonics modules
[49:25] into this industry indry since
[49:28] 2015 um you know we've seen the need for
[49:32] both Dr and fr and at every speed note
[49:36] and so that that was our history when we
[49:38] were a pluggable transceiver purveyor
[49:41] out of Intel we no longer doing that but
[49:43] we are partnering with plugable module
[49:45] vendors and we're providing the key
[49:47] chipsets that go into it and we continue
[49:49] to see the need for both Dr andf FR and
[49:52] at every speed node so as Katherine
[49:55] mentioned this is the front end of the
[49:56] net work um mainly uh but some of these
[50:00] are also making their way in into some
[50:02] backend applications um from what we can
[50:05] tell from our customers and our
[50:06] customers customers so the the
[50:09] transceiver space has clearly shown a
[50:11] need for both uh what we have done when
[50:14] it comes to Optical IO connectivity at
[50:16] Intel uh which we call some of our oci
[50:19] projects and we did a recent
[50:21] demonstration um of this at uh um our
[50:25] our Foundry Day is um a 2 terbit per
[50:28] second package 4 terbit per second total
[50:31] bandwidth by directional and it used
[50:33] eight fibers and eight wavelengths for
[50:36] fiber so we were using both levers uh to
[50:39] cram in that density and in our case
[50:41] that that wavelength Matrix is actually
[50:44] a closely spaced grid um closely spaced
[50:47] wavelengths on a floating grid uh with
[50:49] the associated control technology so we
[50:52] believe that that lever allows you to
[50:55] get up the density
[50:57] um as well as the um the cost per bit
[51:01] and um and the power per bit in line
[51:04] with some of those targets that
[51:06] Katherine had shared earlier although
[51:07] Katherine in your slide you mentioned uh
[51:10] uh 50 cents a bit and I'd love to sign
[51:13] up for that kind of pricing model
[51:16] um so um but but uh but kidding aside I
[51:21] on the next slide when we talk about
[51:22] what the application needs um you know
[51:27] the the power and cost is front of mind
[51:30] um and um and I think you've seen some
[51:32] of these numbers before uh in in the
[51:35] slides from from the panelists um but
[51:38] one one thing that um Katherine uh
[51:41] requested me to focus on and this is
[51:43] something that Dan opened with is also
[51:45] the reliability expectations uh of this
[51:48] as as you're attached to more more and
[51:51] more precious end points and you have
[51:54] more bandwidth flowing to through your
[51:57] particular component the reliability
[51:59] expectations go up and as those
[52:02] reliability expectations go up many of
[52:04] the solutions you see have a very high
[52:06] component
[52:08] count and what that means is the
[52:10] reliability expectation for a given
[52:13] component goes up even higher than that
[52:15] um so so you know we've we've addressed
[52:17] that through some some uh very
[52:19] innovative ways at Intel in addition to
[52:22] many of the other vectors and in my talk
[52:24] I'll talk about that a little bit more
[52:27] so if we go to that next
[52:29] slide one of the things we wanted to
[52:31] indicate here is and this is why when we
[52:34] look at AI applications we think of it
[52:36] as an an or more like a sequence of
[52:39] which levers to pull when as opposed to
[52:42] do I pull just one lever and which one
[52:44] is it um and it's because if you look at
[52:46] the interaction of these as you go up in
[52:48] wavelength count we do see it being
[52:51] beneficial uh to power reduction we do
[52:54] see it being beneficial to cost
[52:56] um but it does have an impact on
[52:59] Shoreline density and overall density of
[53:01] the solution um line rate um you know it
[53:06] at least up to these rates seems to be
[53:08] good across the board and so for those
[53:10] reasons you know uh we're supporting the
[53:12] highest line rate that our technology
[53:14] can support and you know so far at least
[53:17] within oci we've been able to keep up um
[53:21] uh fiber count is in our experience been
[53:24] a weaker modulator of these parameters
[53:28] um the polarization knob is available I
[53:30] think Peter talked about it earlier so
[53:32] this is something that is available in
[53:34] reserve for you know pulling that
[53:37] pulling that lever when when the total
[53:39] bandwidth will actually need it higher
[53:41] order signaling uh going to potentially
[53:44] coherent um again comes with a huge
[53:46] complexity and cost penalty uh but in
[53:49] the front end Network as Katherine
[53:50] mentioned you know we've already pulled
[53:51] that lever right we've gone from nrz to
[53:54] Pam 4 um and even in some of our
[53:57] conversations on oci and CPO in some
[54:00] cases the Shar density needed uh in some
[54:03] of these applications require you to
[54:05] pull that
[54:06] lever so um so our our take on it is
[54:10] these are all knobs that are available
[54:12] uh different applications uh you know
[54:15] Rebecca talked about radex uh you know
[54:17] by definition if you want to go many
[54:20] places the easiest way to do it is with
[54:22] many fibers as opposed to First
[54:24] combining it into a single fiber then
[54:26] splitting it out which has huge penalty
[54:28] so it's very application dependent
[54:30] within our platform uh we support
[54:33] multifiber attached in an automated way
[54:36] like Rebecca mentioned um we have uh
[54:39] both course wdm as well as a finely
[54:42] spaced uh oand wavelength capability on
[54:45] the same chip um so what we're trying to
[54:48] do is make sure that the platform is
[54:50] ready and has these Le levers available
[54:53] to invoke in a cost effective and
[54:55] reliable manner when the application
[54:58] demanded next
[55:00] please and so this is talking a little
[55:03] bit about you know what most people
[55:05] perceive
[55:07] as the probably the single weakest link
[55:10] in an optical link from a reliability
[55:12] perspective I have a slightly different
[55:14] point of view but I want to address the
[55:16] laser anyway since everybody talks about
[55:18] it I think fiber attach is the weakest
[55:20] link um but um uh probably firmware
[55:25] quality is even higher than that but you
[55:27] know um fiber from a physical part
[55:31] perspective is probably the weakest link
[55:33] and goes back to many of the things that
[55:35] Don Dan talked about how how fibers made
[55:38] how they connected you know the manual
[55:39] nature of processes the materials the
[55:42] mismatches but focusing in on labor
[55:44] lasers lasers are very well understood
[55:47] and some of the mechanisms that lead
[55:49] lasers to fail are also very well
[55:51] understood and when we put together this
[55:54] silicon photonics platform with the
[55:56] integrated laser which has the you know
[55:58] 35 material co-processed with our 300 mm
[56:01] silicon Wafers uh we we we leverage that
[56:05] opportunity to actually solve many of
[56:07] the issues that cause laser failure by
[56:09] Design so for example facets are a
[56:12] failure point we don't have facets in
[56:14] our architecture we have eved coupling
[56:16] of light in and out of the gain material
[56:19] so you eliminate a failure mode by
[56:21] category um you know the regrowth
[56:23] interfaces have been eliminated um we
[56:26] put the gratings in the Silicon as
[56:28] opposed to the 35 material um so so uh
[56:32] many of these things where we've seen
[56:35] you know where where lasers have
[56:37] traditionally failed um you know we've
[56:39] addressed them by Design then that still
[56:41] leaves another category which is
[56:43] material quality and random defects
[56:45] typically in photonics we don't care too
[56:47] much about random defects in photonics
[56:49] typically the focuses on parametric
[56:51] yield RR devic is good enough that you
[56:53] can Cascade them together to meet your
[56:55] functional specs and I think we're
[56:57] getting to a point whereas the digital
[56:59] seos world is very different you know
[57:01] parametric yields are insanely high and
[57:04] people focus about on defect density and
[57:06] as a result of that 300 millimeter
[57:08] manufacturing lines especially digital
[57:10] seos lines are very well suited to track
[57:13] identify defects and then find out what
[57:16] is their root cause and eliminate them
[57:18] to process control and these are
[57:20] methodologies that we've embraced and
[57:22] while for the rest of the photonic
[57:24] circuit it doesn't matter too much
[57:25] because it's very SPS this has led to a
[57:28] lot of improvement on the laser side um
[57:31] as a result of that our field experience
[57:34] as well as our Rel internal reliability
[57:37] programs where we run a huge number of
[57:38] volumes through uh stress testing our
[57:41] experience has been laser fit rates of
[57:44] less than 0.1 per device and uh for the
[57:47] rest of the industry low single digits
[57:50] is an aspirational goal uh we didn't
[57:52] lock into it it was partly by Design and
[57:54] a lot of uh hard hard fought battles it
[57:57] hasn't mattered much in the transceiver
[58:00] space u because you don't have the
[58:02] device count and like I mentioned
[58:03] there's other things that fail at a much
[58:05] higher rate than the lasers do uh but
[58:08] for this next Frontier we're ready and
[58:11] excited next slide please the quick
[58:14] recap we don't believe uh one approach
[58:17] you know all multifiber or multi Lambda
[58:19] fits all applications you will need
[58:22] multiple multiple levers um we plan to
[58:26] continue to support within the frontend
[58:28] network uh plugable components you know
[58:31] both the Dr and fr formats and when it
[58:34] comes to the uh connectivity Market the
[58:36] xpu connectivity Market our solution
[58:38] relies on both it relies on both today
[58:41] and we plan to keep relying on both as
[58:43] we go
[58:45] forward thank you thank you very much
[58:48] Amit that was uh very very thorough and
[58:51] thank you for digging into the topic of
[58:53] reliability it's a very important one
[58:56] now I don't see any questions and I
[58:59] don't have any clarifying questions Dan
[59:01] did you have anything top of
[59:05] mind if not I think in the interest I
[59:08] was muted um just on this slide see I
[59:10] have access to the slides and uh in the
[59:12] bottom right you're showing an indium
[59:14] phosphide die on a silicon wafer um how
[59:17] do you do that integration yeah so it's
[59:20] very interesting um there's many ways of
[59:22] hybrid integration and often what people
[59:24] will do find a way to process your
[59:27] silicon wafer process the Indian
[59:28] phosphide die coupon eyesee or you know
[59:32] segregate the Indian phosphide and then
[59:33] align or integrate the way we do it is
[59:36] we do the front section of the device we
[59:38] do wave guides and all kinds of things
[59:40] gratings early on on a 300 mm wafer and
[59:44] then uh we have an optimized Epi that we
[59:47] do in house which is designed for our
[59:49] integration um we take those three five
[59:52] coupons uh we take those 3in Wafers of
[59:55] Indian phospide we coupon eyesee them
[59:58] and we flip chip Bond them uh and mass
[01:00:01] onto a huge wafer 300 mm wafer and then
[01:00:04] we co-process them so it's not a laser
[01:00:06] it's an Epi stack before you put it on
[01:00:10] and you co-process the laser at 300 mm
[01:00:13] scale um so so it's a you know
[01:00:17] physically aligned that's how we can get
[01:00:19] light in and out through essent coupling
[01:00:21] but uh so it's not an active alignment
[01:00:23] of a laser after the fact right thank
[01:00:26] you so much very very clever thank you
[01:00:29] um thank you very much amid I would like
[01:00:32] to call our next speaker forward Alan
[01:00:35] Lou Alan Lou is um CEO and co-founder of
[01:00:41] quintessent and uh you may have seen him
[01:00:44] he's been active in the industry for
[01:00:45] some time if you've worked on a Dara
[01:00:47] program you have probably seen Alan
[01:00:49] there so please go ahead Alan tell us
[01:00:52] about quintessent and your
[01:00:54] solution sure like Dan is still pulling
[01:00:57] up the slides oh here we go um so um
[01:01:01] well first of all thanks Dan and
[01:01:02] Katherine for the invitation it's a
[01:01:03] honored to kind of be here with the rest
[01:01:05] of the Steam panelist and looking
[01:01:06] forward to the discussion and I'm
[01:01:08] learning a lot um so at quintessence
[01:01:10] we've been working on high W accounts
[01:01:12] specifically dwdm Technologies for uh
[01:01:15] dwdm based interconnect architectures
[01:01:17] and so with respect to the panel
[01:01:20] discussion topic um you know I have a
[01:01:22] very clear opinion which is that life is
[01:01:25] better in color and for the avoidance of
[01:01:28] doubts and ambiguity by color I mean
[01:01:30] wavelengths so next slide please Dan and
[01:01:34] I I'll share my
[01:01:35] perspective um from the I'll kind of
[01:01:39] examine the question from the lens of
[01:01:41] you know both performance and
[01:01:42] operational considerations and obviously
[01:01:45] a nuanced topic but in the interest of
[01:01:47] time I'm GNA try to distill you know it
[01:01:50] down to sort of what I perceive as the
[01:01:53] key considerations so first of all why
[01:01:55] not not more fiber obviously AI clusters
[01:01:58] today um use a lot of breakout
[01:02:01] applications for increas in radex dr4
[01:02:03] dr8 sort of a predominant um Workhorse
[01:02:06] for the backend GP to GPU connection um
[01:02:10] but in terms of looking at the scaling
[01:02:12] on in the future you know one important
[01:02:14] metric is that's um not only is it
[01:02:16] important to have radex but bandwidth
[01:02:18] per fiber or bandwidth per effective
[01:02:20] radex ports also needs to grow and so
[01:02:23] that by definition eliminates scaling
[01:02:27] just with adding more fibers alone in
[01:02:30] parallel um uh or you know takes that
[01:02:33] option off the table in terms of trying
[01:02:35] to meet this requirement of
[01:02:37] accommodating bandwidth per fiber growth
[01:02:40] uh next
[01:02:41] please the other consideration is that
[01:02:43] um reliability at a system scale at the
[01:02:46] data center scale and um many of the
[01:02:49] panelists before me have touched on this
[01:02:51] before there's obviously lots of
[01:02:53] different considerations here in terms
[01:02:55] of what what determines the overall
[01:02:56] system level reliability um starting
[01:02:59] from the material Level to the component
[01:03:01] to the link and then all the way to the
[01:03:02] application Level but we kind of look at
[01:03:05] you know the key relevant ingredients in
[01:03:08] the context of the discussion um every
[01:03:11] mechanical moving Parts in the data
[01:03:14] center or in the cluster is a potential
[01:03:17] point of failure every fiber termination
[01:03:20] has potential to have dust in it um
[01:03:23] every um uh every additional a fiber has
[01:03:26] potential to Break um etc etc and also
[01:03:30] um every additional parallel fiber has
[01:03:33] adds complexity and cost in terms of
[01:03:35] alignment um and also present some eeld
[01:03:38] issues there so trading off fiber
[01:03:40] terminations um where you can for things
[01:03:43] that can be brought onto the chip on the
[01:03:46] wafer um is the path to improving
[01:03:49] overall system level reliability because
[01:03:50] that eliminates you know it moves
[01:03:52] failure points from things that are
[01:03:53] mechanical and moving to things that are
[01:03:56] stable and on the wafer next slide
[01:03:59] please or next click Dan um and then
[01:04:02] finally you know I'll make the notes
[01:04:04] that uh cabling in state-of-the-art
[01:04:07] large scale GPU clusters like the one
[01:04:10] that was just recently brought up by
[01:04:12] Tesla and x. is already a nightmare so
[01:04:14] this photo is you know showing the Sea
[01:04:18] of yellow SMF fibers that are prevalent
[01:04:20] and also just bundled on the floor and
[01:04:24] uh overall that looks like a a messy
[01:04:28] Nightmare and if we're asking whether we
[01:04:30] want to add to that mess pretty scary
[01:04:33] proposition to me so um in addition to
[01:04:36] this you know I think another
[01:04:37] consideration is that obviously for
[01:04:39] Green Field deployments of GPU clusters
[01:04:41] you have the flexibility to lay down you
[01:04:43] know brand new fiber plants and fiber
[01:04:45] architecture and installations but I
[01:04:47] think in general uh in the past and
[01:04:49] likely in a desire and as a desire for
[01:04:51] the future um it would be nice to put
[01:04:54] down a five architecture and uh fiber
[01:04:57] plant on day one and I have that last
[01:04:59] across many multiple GPU generations and
[01:05:02] so from that perspective it's nice to be
[01:05:04] able to reuse the existing fiber
[01:05:07] infrastructure when you're going through
[01:05:09] multiple upgrade cycles and scaling your
[01:05:11] wavelengths by just adding more
[01:05:12] wavelengths per fiber and reusing the
[01:05:14] fiber infrastructure um allows you to do
[01:05:16] that whereas adding in more strands of
[01:05:19] fiber is a huge operational um
[01:05:23] undertaking so next slide please Dan
[01:05:26] all right so let's play The dev's
[01:05:28] Advocate for why not more wavelengths um
[01:05:31] you know one thing that I think uh I see
[01:05:34] around the industry is that um you know
[01:05:36] in terms of the choice of um one
[01:05:39] potential consideration is just you know
[01:05:42] fonic elements especially components on
[01:05:44] the chip usually have a finite um
[01:05:48] spectral band whether we response and
[01:05:51] specifically in the context of the
[01:05:53] choice of fiber couplers there's really
[01:05:55] only two choices Edge coupling and
[01:05:57] vertical coupling um and each one
[01:06:01] everyone has their religion between the
[01:06:02] two um Some people prefer vertical
[01:06:05] coupling for high through putut wafer
[01:06:08] level testing and Manufacturing and in
[01:06:10] those cases um where that is a strong
[01:06:14] consideration then you would be
[01:06:16] limitating your flexibility of um or in
[01:06:19] the total range of the wavelength
[01:06:21] Spectra that you have to play with for
[01:06:22] example you want to be able to use a
[01:06:24] cwdm grid and you want to able to just
[01:06:26] add um additional course wavelengths um
[01:06:29] at well for example but one solution to
[01:06:32] get around that um is to move to a
[01:06:34] denser grid so instead of 20 nanometer
[01:06:37] cwdm um maybe consider 2 nanometer dwdm
[01:06:41] for example and that's what we're
[01:06:42] interested in in exploring AC
[01:06:45] quintessence uh next click please Dan um
[01:06:48] the other
[01:06:49] traditional I guess arguments against um
[01:06:52] wavelength scaling which has been
[01:06:53] touched on uh in this panel already
[01:06:55] that's wavelength scaling is expensive
[01:06:58] because um traditionally when the method
[01:07:02] of scaling Band withd by adding
[01:07:04] wavelengths is requires the addition of
[01:07:07] an additional laser the incremental cost
[01:07:11] of that wavelength or exual bandwidth is
[01:07:14] the cost of another laser and so you
[01:07:15] have a pretty steep um incremental cost
[01:07:19] of bandwidth scaling curve that is more
[01:07:21] or less the cost of adding more lasers
[01:07:23] uh or for additional waveling per fiber
[01:07:26] but one solution to get around this is
[01:07:29] um again if you're using a dwdm GD um
[01:07:32] Advance please Dan um then you have the
[01:07:35] option to leverage um you know advances
[01:07:38] in comb Blazers specifically dwm comb
[01:07:41] Blazers that gives you the functionality
[01:07:43] and wavelengths of multiple lasers but
[01:07:46] for the price of one and so that breaks
[01:07:47] this cost scaling curve where instead of
[01:07:50] the Steep um increasing cost as you add
[01:07:53] more wavelength to per fiber because you
[01:07:55] have to pay for an additional laser um
[01:07:57] that cost can be pretty flat because
[01:07:59] you're advertising the cost of a single
[01:08:00] laser over multiple wavelengths the
[01:08:03] third um I guess uh argument against
[01:08:06] wavelengths um next click thank you Dan
[01:08:09] is that uh wavelengths are difficult to
[01:08:11] control and again um I would say that
[01:08:14] has been true traditionally when you're
[01:08:17] architecting and building interconnects
[01:08:19] that use single lasers per Lambda um but
[01:08:24] again there's Solutions do this so Dan
[01:08:26] if you can advance please um if you're
[01:08:29] already using the dwdm grid and you're
[01:08:32] using dwdm comb lasers to break the cost
[01:08:35] scaling curve um then you can also
[01:08:37] leverage the fact that um you know at
[01:08:39] least for the lasers that quintes makes
[01:08:42] um the wavelength spacing between
[01:08:45] wavelength frequencies and the com laser
[01:08:47] is very consistent and very repeatable
[01:08:49] and so if you know where one wavelength
[01:08:51] is you know where the neighbors are and
[01:08:53] all of those wavelengths track and drift
[01:08:56] together and so that that's a lever that
[01:08:58] allows you to then simplify the control
[01:09:02] of the wavelengths in the high W length
[01:09:04] account dwdm link that uses this
[01:09:06] technology and we've shown that it's
[01:09:08] very you can get very repeatable and
[01:09:10] very predictable spacings in comb lasers
[01:09:12] for the frequency variation is less than
[01:09:15] 2% um so next slide please
[01:09:19] Dan so in the final analysis um you know
[01:09:23] another observation is that uh the two
[01:09:26] aren't necessarily mutually
[01:09:28] exclusive uh you can scale up uh in
[01:09:32] fibers to achieve more radic um and on
[01:09:36] top of that that you can also then add
[01:09:37] more wavelengths to the same fiber so um
[01:09:40] both can be used to you know depending
[01:09:43] on the application that need and so for
[01:09:45] example whereas 800g dr8 is sort of the
[01:09:48] Workhorse interconnects um for um
[01:09:52] backend networks today um before we
[01:09:55] think about adding more fibers um why
[01:09:57] not shove more wavelengths down each
[01:09:58] fiber and so you know potentially we can
[01:10:01] have eight wavelengths per fiber stay on
[01:10:03] a Dr configuration for the radex but now
[01:10:05] we're talking 6.4 terabits total or 800g
[01:10:09] per fiber which is very interesting so
[01:10:11] in general I would say reliability um
[01:10:15] Simplicity of deployment and upgrades um
[01:10:18] and also the need to drive up per fiber
[01:10:21] bandwidth growth necessitates wavelength
[01:10:23] scaling um obviously radex is very
[01:10:25] important too which has been talked
[01:10:26] about in this panel and um that favors
[01:10:30] having more fibers for the breakout um
[01:10:33] but you know that up to a limit right
[01:10:36] you know you only have certain uh a
[01:10:38] certain number of ports on your switch
[01:10:41] and so if you have more breakouts than
[01:10:44] the port count then you're ask you're
[01:10:46] talking about either adding more
[01:10:47] switches or more layers of the switch
[01:10:49] which then becomes another cost driver
[01:10:52] um so in the final analysis uh Dan if
[01:10:55] you can click one more time um my
[01:10:58] assessment is that both are likely
[01:10:59] needed to scale up AI clusters um but
[01:11:02] obviously in my opinion life is always
[01:11:04] better with more color all right thanks
[01:11:07] very much thank you very much Alan and I
[01:11:10] have a quick question for you again
[01:11:12] clarifying how many wavelengths are we
[01:11:15] talking how how many times can you
[01:11:19] double we see a pretty good path to
[01:11:21] having 32 wavelengths on a on per per f
[01:11:25] and that's what we're working on or
[01:11:27] towards um initially it'll be in groups
[01:11:30] of eight um and you know Peter talked
[01:11:33] about that um uh it's not ideal to run
[01:11:37] each wavelength slow um and I'll just
[01:11:40] add that wavelength scaling and wdm
[01:11:42] doesn't and wide and parallel doesn't
[01:11:45] have to mean wide and slow and parallel
[01:11:47] you can have lots of wavelengths and
[01:11:48] still run them a modest speeds right so
[01:11:51] um in this Dr example I'm imagion that
[01:11:54] each wavs at 100 Gig per
[01:11:57] thank you well thank you very much Alan
[01:12:00] thank you and I don't see any questions
[01:12:02] for Allan here on any of the channels so
[01:12:06] I will go ahead and call our final
[01:12:07] speaker bardier would you come to the
[01:12:10] stage and
[01:12:14] um I don't see viia coming yet somebody
[01:12:17] could help
[01:12:20] vardia so let me go ahead and introduce
[01:12:22] them there's enough people backstage I'm
[01:12:24] sure sure who can help him come through
[01:12:27] um so bardier is uh co-founder and CEO
[01:12:32] of
[01:12:33] avisena and um Bia is a Serial
[01:12:36] entrepreneur so he this is not his first
[01:12:39] rodeo he's done many startups in the
[01:12:41] past I would like to hear Bia tell us
[01:12:43] about um why you think this solution is
[01:12:46] the
[01:12:47] winner yeah uh it's great to be among
[01:12:50] such AUST company um going to talk about
[01:12:54] something quite different
[01:12:55] not laser based but using LEDs that I
[01:12:58] think really solve some of the problems
[01:13:00] we've been struggling with maybe the
[01:13:02] next
[01:13:04] slide so we've discussed in depth how AI
[01:13:09] clusters and um generally ic's have a
[01:13:12] copper problem and I think Katherine
[01:13:15] alluded to this picture in her
[01:13:17] presentation where the Nvidia system has
[01:13:20] 5,000 copper cables and these cables are
[01:13:23] sucking enormous amounts of power
[01:13:25] and they limit the number of gpus that
[01:13:28] can go around the switch Jensen famously
[01:13:31] in his speech declared very proudly that
[01:13:34] he's not using Optics and he's very
[01:13:36] happy about not using Optics and the
[01:13:38] factors that were mentioned was that the
[01:13:40] Optics is unreliable it's expensive it's
[01:13:43] bulky and it consumes a lot of power so
[01:13:46] it's stuck to Copper but issue with
[01:13:49] copper is that you just can't put more
[01:13:51] gpus around that switch this copper
[01:13:54] problem extends everywhere I think we
[01:13:55] all know how processing power in chips
[01:13:58] has increased because of Moors law but
[01:14:00] iio really hasn't mors law hasn't done
[01:14:03] much for the resistance of copper or the
[01:14:05] capacitance of free
[01:14:08] space maybe next
[01:14:11] slide okay so we think that the problem
[01:14:14] is the fiber optics that we're using is
[01:14:17] really derived from technology that was
[01:14:19] developed for long Holdings so this is
[01:14:22] fiber optics lasers and single mode
[01:14:25] fiber or great if you want to go across
[01:14:26] the oceans if you want to go across the
[01:14:28] continent but if you want to go to the
[01:14:30] grocery store go inside the Rack or go
[01:14:33] you know a few meters it's the wrong
[01:14:35] technology so making you know using that
[01:14:37] technology that was developed
[01:14:39] fundamentally for long distances and
[01:14:40] applying it for short distances means
[01:14:43] you've got a a lot of pain points you
[01:14:45] have to get over just a simple example
[01:14:48] the wavelengths that we choose 1.55
[01:14:51] microns or 1.3 microns that's the fiber
[01:14:54] the glass glass minimum loss and glass
[01:14:56] minimum dispersion wavelengths uh we use
[01:14:59] those because we want the the signals to
[01:15:02] go far well if you're going a short
[01:15:04] distance you don't really care that the
[01:15:05] loss is a qued DB per kilometer right uh
[01:15:08] you can tolerate to have more more loss
[01:15:11] so we' like to talk about a different
[01:15:13] approach that can address those problems
[01:15:15] so next
[01:15:17] slide and instead of piggybacking off
[01:15:19] the fiber optics World we're instead
[01:15:22] piggybacking from the display world
[01:15:25] uh it turns out that people as smart as
[01:15:28] us have been working in a whole
[01:15:29] different area which is displays and a
[01:15:32] technology has been developed for
[01:15:35] putting millions of LEDs on Silicon
[01:15:39] chips um Apple was going to use this
[01:15:42] microed display for watches I think they
[01:15:45] recently actually cancelled that program
[01:15:47] sticking with oleds but all the big
[01:15:50] players are looking at these types of
[01:15:52] displays for virtual reality for glasses
[01:15:55] they're actually in production for car
[01:15:56] headlights as's a picture of that sort
[01:15:59] of in the middle and you can see all
[01:16:00] these teeny little LEDs Gallum nitrite
[01:16:02] LEDs on Silicon chips there was one
[01:16:05] company in the area here Mojo that had
[01:16:08] 100,000 pixel display they put in a
[01:16:10] contact lens so the maturity of this
[01:16:13] technology and what is capable of is
[01:16:15] amazing now normally people don't think
[01:16:17] about using LEDs for communications
[01:16:20] fundamentally because traditionally LEDs
[01:16:22] have been slow you can't turn them on
[01:16:25] off very fast you're relying on
[01:16:27] spontaneous emission versus stimulated
[01:16:29] emission um however it turns out next
[01:16:32] slide that the gallium nitride LEDs that
[01:16:36] are used for displays can be tweaked to
[01:16:39] run extremely fast uh fundamentally this
[01:16:42] is because electrons and holes in Gallum
[01:16:44] nitride really like each other and you
[01:16:46] can use that to make a very efficient
[01:16:48] LED or you can use the same behavior to
[01:16:51] make a very fast LED I think this was
[01:16:53] about a year ago we did a live demo of
[01:16:56] our LED Link at a at a conference and we
[01:16:59] were modulating this led on and off of
[01:17:01] 14 gbits per second and we put arrays of
[01:17:05] these LEDs that's the lower picture on
[01:17:08] Silicon chips and when you have arrays
[01:17:11] of hundreds thousands possibly even you
[01:17:14] know millions of LEDs all modulating at
[01:17:17] these speeds you can see that the
[01:17:18] overall bandwidth can be incredible um I
[01:17:22] think one of the presenters uh uh
[01:17:25] mentioned how clock speeds on chips
[01:17:27] hasn't changed in in decades and clock
[01:17:30] and the chips themselves move data
[01:17:32] relatively slow rates these LEDs are
[01:17:35] very compatible with these types of
[01:17:37] speeds and they can scale to the kind of
[01:17:39] numbers that you need to get data on and
[01:17:41] off chips um the picture down below is
[01:17:45] our little demo chip and sometimes
[01:17:46] there's a video that plays I don't know
[01:17:48] if this one does but the LEDs turn on
[01:17:50] and off and they send data in and out of
[01:17:52] that chip so this was a 16 nanometer
[01:17:55] tsmc chip it had 300 LEDs uh each LED
[01:18:00] was running about 3 to four gbits per
[01:18:02] second so sending about a terabit of
[01:18:04] data and doing it about a peer jeel per
[01:18:06] bit the advantage of using LEDs versus
[01:18:09] lasers is is tremendous you know in a
[01:18:12] laser you need to drive enough current
[01:18:14] to get it over threshold you know just
[01:18:16] like an airplane you got to get it
[01:18:18] Airborne first before anything happens
[01:18:20] um LEDs there's no threshold you're
[01:18:22] moving around on the ground you don't
[01:18:24] care about having isolators if this feed
[01:18:27] back into the laser there's really no
[01:18:29] polarization issues there's really no
[01:18:31] modes of uh to speak of frequently with
[01:18:35] lasers we have to be very careful to AR
[01:18:37] code everything and pay attention to our
[01:18:39] coupling because you can get bit error
[01:18:41] rate floors and this really comes
[01:18:43] because you're talking about a narrow
[01:18:45] line width very long coherence length
[01:18:47] basically it's speckled uh with LEDs
[01:18:50] there is no speckle so you don't get
[01:18:52] these types of bit err rate floors these
[01:18:54] LEDs can work great at high temperatures
[01:18:56] I think recently some friends of ours
[01:18:58] published the paper on our devices and
[01:19:00] they took it up to 400° C operation so
[01:19:03] your silicon is going to die before your
[01:19:05] Gallum nitride LED dies we have
[01:19:07] extremely good reliability incredibly
[01:19:10] low cost these millions of LED displays
[01:19:14] are selling for tens of dollars so the
[01:19:16] cost of the LEDs is extremely low as you
[01:19:19] know Gallum nitrite is used in LED
[01:19:21] lighting Humanity makes almost square
[01:19:24] miles the stuff so the cost of the dyes
[01:19:27] are very very low and and as I as I
[01:19:29] mentioned you can have massive
[01:19:31] parallelism so there's no series there
[01:19:34] was quite a bit in this panel about the
[01:19:37] fibers unfortunately I don't have a
[01:19:38] slide on the fibers but we're not using
[01:19:42] standard F Celica fibers that are quite
[01:19:45] expensive uh and again are really
[01:19:47] optimized for super low loss the fibers
[01:19:50] we're using are fibers borosilicate
[01:19:53] fibers from lamp fixture
[01:19:55] so these are used in lighting
[01:19:56] applications that typically come in
[01:19:58] bundles of about 4,000 and they used in
[01:20:01] things like microscope lights and
[01:20:03] sometimes in aircraft or pool lights uh
[01:20:06] so we use those types of fibers we
[01:20:07] bundle them up in this case it's a
[01:20:09] bundle of 300 but you can make bundles
[01:20:11] of of thousands and the diameter
[01:20:14] controll is accurate so you can put them
[01:20:17] in the right position so you can make
[01:20:18] simple connectors uh and cables and we
[01:20:21] have all of those things uh within the
[01:20:23] company
[01:20:25] let's go to the next slide
[01:20:27] maybe um this is the one terabit demo
[01:20:30] chip that we showed at the optical fiber
[01:20:32] conference the two hexagons here are the
[01:20:35] transmitters and the receivers so the
[01:20:37] right hand hexagon is 330 LEDs it's
[01:20:41] basically little TV but it has a frame
[01:20:44] rate of about 3 to four gigabits per
[01:20:47] second so three billion frames a second
[01:20:50] versus 60 frames a second um the other
[01:20:52] side of it is the camera um in this case
[01:20:55] we didn't yet have access to camera
[01:20:58] technology from tsmc and the other
[01:21:00] foundaries so we actually made our own
[01:21:02] little camera it's a array of 300
[01:21:05] silicon uh photo detectors made on a
[01:21:08] separate ship and we turned it upside
[01:21:10] down and bonded on top of the main tsmc
[01:21:13] chip in some processes you can make the
[01:21:16] detectors in the chip itself in this
[01:21:18] particular process we can figure out how
[01:21:19] to do that so we have a separate uh
[01:21:22] photo detecting layer stuck on top top
[01:21:24] of it uh we use blue light and blue
[01:21:28] light is absorbed very easily in Silicon
[01:21:30] so you can make very efficient highspeed
[01:21:33] low capacitance per unit area types of
[01:21:36] detectors that are easy to couple to and
[01:21:38] can be Amplified uh with very low energy
[01:21:41] this chip had a lot of functions in it
[01:21:43] it had its own birt and PPG it had a eye
[01:21:46] Monitor and it's a it's a it's a pretty
[01:21:49] cool chip
[01:21:52] um next slide Maybe
[01:21:56] where are we going with this technology
[01:21:58] it's that co-packaged Optics idea that
[01:22:01] Katherine mentioned against in the
[01:22:02] beginning of the presentation and I
[01:22:04] think uh um other guys have referred to
[01:22:07] this is implementation of our technology
[01:22:10] as a chiplet where we make a 12 terbit
[01:22:12] chiplet and you can see we can fit it
[01:22:14] into a 5 mimer Shoreline uh in this case
[01:22:18] we using we're going to be using a UCI
[01:22:21] interface so the shoreline dens is
[01:22:24] actually limited by the
[01:22:26] ucie and we double the speed of our LEDs
[01:22:29] from 4 gig to 8 gig as you seen we've
[01:22:32] gone up to 14 gig in the past and we
[01:22:35] connect a larger array of fibers onto
[01:22:37] this device uh so that's where we think
[01:22:40] the technology is going to go from the
[01:22:42] demo chip to a co-packaged Optics and
[01:22:45] then what's really exciting about this
[01:22:47] is you can put these LEDs and detectors
[01:22:49] really on any chip it doesn't even have
[01:22:51] to be a separate chiplet it could go on
[01:22:54] the GPU ultimately or on the memory uh
[01:22:57] hbm usually has an active chip under the
[01:23:00] memory stack so you could go into that
[01:23:02] active chip underneath and be able to
[01:23:04] connect directly to the bus that's
[01:23:07] running on the GPU and eliminate the
[01:23:10] seris next
[01:23:14] slide so just to summarize U people have
[01:23:18] dreamt about putting Optics on chips for
[01:23:20] years people complaining that silicon
[01:23:22] doesn't give off light well you put
[01:23:23] little micro LEDs on Silicon then
[01:23:25] obviously silicon does give off light uh
[01:23:28] and you can make very large arrays of
[01:23:30] these LEDs this technology is no good
[01:23:33] for long distances this is not a
[01:23:34] technology you want to drive across the
[01:23:36] country um it but it's fantastic for
[01:23:40] centimeters to meter type scale it's
[01:23:42] simple it's low cost it goes on any
[01:23:45] silicon note reliable gives you the wide
[01:23:48] buses and it's a right solution to
[01:23:51] Jensen's problem of all those 5,000 C
[01:23:56] wires thank you thank you very much B
[01:24:00] very that was a very interesting
[01:24:01] discussion quite different to what we've
[01:24:03] heard from the other speakers um maybe
[01:24:05] if I can just ask you one really quick
[01:24:07] question so you heard Alan ask about
[01:24:10] upgradability and particular adding new
[01:24:13] fiber and um you've made a point here
[01:24:16] that this is a very short distance so
[01:24:18] does that give you another degree of
[01:24:20] Freedom how would you upgrade to double
[01:24:22] your fiber or double your bandwidth
[01:24:25] yeah um so you can obviously increase
[01:24:28] the bandwidth by increasing the speed of
[01:24:30] the LEDs our first demo chips are 4 gig
[01:24:32] we're going to go to 8 gig that we use
[01:24:34] the same fibers but frankly when the
[01:24:36] fibers are relatively short and they're
[01:24:38] relatively inexpensive it's not that
[01:24:40] hard to replace that fiber um currently
[01:24:44] we have 50 Micron diameter fibers which
[01:24:48] sets the spacing of the LEDs these
[01:24:51] fibers also go uh come in at 25 micron
[01:24:54] diameter and you can even go down to 15
[01:24:57] and if you multiply those out go to 10
[01:24:59] gig per LED and go to 15 Micron diameter
[01:25:03] fibers of spacing you're at almost 10
[01:25:06] terabits per millimeter squared so very
[01:25:09] very high densities are achievable with
[01:25:11] this technology and of course the number
[01:25:13] of LEDs is almost unlimited you know
[01:25:15] displays run on millions of
[01:25:18] LEDs
[01:25:19] yeah thank you very much so we're only
[01:25:22] limited by the patients of our audience
[01:25:25] and I'd say we can certainly take 20
[01:25:26] minutes if we have questions to fill
[01:25:28] that and then we'll see how it goes
[01:25:30] excellent um and I'd like to suggest um
[01:25:33] Dan and I Dan Pitt and I prepared a
[01:25:36] little Poll for the audience and I'm
[01:25:39] thinking maybe we start that now and uh
[01:25:42] we have the audience think about how
[01:25:44] they would vote so the question um Dan I
[01:25:48] can launch this from here I
[01:25:51] think I thought I could polls and
[01:25:53] quizzes
[01:25:55] it's Dan's poll I'm launching
[01:25:58] it okay and so the audience should be
[01:26:01] able to see it on my screen it's a popup
[01:26:04] that came up and you are able to vote
[01:26:08] and I see real time voting happening as
[01:26:10] we watch this is
[01:26:11] amazing uh so we'll let this run for a
[01:26:14] few minutes here while we ask questions
[01:26:16] so I'd like the audience to answer the
[01:26:19] question what's the best way to scale is
[01:26:21] it wavelength fiber or you Underside
[01:26:25] and within the undecided there's
[01:26:26] probably it depends because there's
[01:26:29] probably some in it depends as well
[01:26:31] let's see how this goes and maybe um
[01:26:34] we'll let it run for a little while see
[01:26:35] if we are still adding votes and um ask
[01:26:39] some questions so I'd like to get a
[01:26:41] sense for how much can we scale how many
[01:26:45] how what's the biggest bandwidth that we
[01:26:47] can do by either of these approaches so
[01:26:49] Bia you just shared that you could do
[01:26:52] 12T um terabits per second I heard uh
[01:26:57] Allan say 32 wavelengths and maybe we
[01:27:01] try and get this all in the same units I
[01:27:02] think ter terabits per second is
[01:27:04] probably a common language for both
[01:27:06] approaches so if this barter is has has
[01:27:09] started the bidding at
[01:27:11] 12T um then what is the bandwidth using
[01:27:15] wavelengths how how much would be
[01:27:17] practical to to put if you've got 32
[01:27:20] wavelengths and Alan I assume question
[01:27:24] directed at me so go for it so so
[01:27:27] breakout fibers are running at 200 gig
[01:27:29] per Lambda today so there's lots of room
[01:27:31] to go and um our road map we see
[01:27:34] basically headro up to four terabits per
[01:27:36] fiber using 32 wavings so plenty of room
[01:27:39] from where we are
[01:27:41] today so I heard four terabits yes per
[01:27:45] fiber that's cor denominator here
[01:27:49] terabits per millimeter or terabits per
[01:27:52] something right because obviously then
[01:27:54] you can scale into whatever you want uh
[01:27:57] with big area we we this is going to get
[01:27:59] very complicated I was actually just
[01:28:01] thinking of not doing it per millimeter
[01:28:03] but just doing you know what's the
[01:28:05] bandwidth so terabits per second and but
[01:28:10] you're right SI size is important so you
[01:28:13] know it's got to fit together with a
[01:28:15] chip probably in a in a relatively
[01:28:17] limited space Shoreline density is
[01:28:19] always limited um so bardia you have
[01:28:23] kind of a special cas Cas with the fiber
[01:28:25] there because you're using um
[01:28:27] illumination fiber how about one of the
[01:28:30] panelists who is more focused on single
[01:28:34] mode um single mode Fibber Solutions and
[01:28:37] putting multiple fibers maybe um either
[01:28:40] Peter or Rebecca Peter do you have an
[01:28:43] opinion about how many fibers you could
[01:28:47] do so I mean as as Dan also mentioned
[01:28:50] it's all a yield question so how much
[01:28:53] how much can you yield on your on your
[01:28:55] solution 16 full duplex channels is a is
[01:28:58] a good choice 32 uh okay and 32 time 200
[01:29:03] gig uh that gives you uh your 12.8
[01:29:07] terabits per module uh TX plus RX
[01:29:11] counting TX plus RX as is typical in
[01:29:13] this uh Community um as well so that's
[01:29:17] totally doable so what was the number
[01:29:19] you
[01:29:20] said 12.8 Tera 12.8 okay
[01:29:25] Rebecca uh yeah I mean I I agree to
[01:29:28] first order uh what Peter said um I
[01:29:31] think the way to look at it is how is
[01:29:34] technology scaling uh for things like
[01:29:37] switches and xpu needs and do we match
[01:29:42] the density of their Shoreline off the
[01:29:44] die that they could get and if you can
[01:29:46] do that then you're good right so 12.8
[01:29:50] terabits is is great but without a
[01:29:53] number to scale that by and millimeter
[01:29:55] Shoreline it's actually relatively
[01:29:58] meaningless um so I would look at
[01:30:00] technologies that are pushing those
[01:30:02] boundaries so things like the switch
[01:30:04] technology is pushing those boundaries
[01:30:07] pretty dramatically um where you have
[01:30:10] you know something like 512 IO out of
[01:30:15] the entire uh perimeter of the chip
[01:30:18] right so 128 IO out of a retical edge
[01:30:22] that is the density that you need to get
[01:30:25] toward and ensuring that you have a
[01:30:27] pathway to scale with that technology is
[01:30:30] what's
[01:30:31] important thank you I'm going to ask
[01:30:34] some more of you for this this this see
[01:30:36] if there's any more bidding on on how
[01:30:39] many terabits per second we can get to
[01:30:41] here but uh just to clarify somebody um
[01:30:44] has asked me in the poll that we're
[01:30:47] voting on why wasn't there an option of
[01:30:49] both uh just trying to keep things
[01:30:52] simple so I think both would be
[01:30:54] undecided because it could be either so
[01:30:56] undecided either or both it depends all
[01:31:00] of that that's all undecided I'm looking
[01:31:03] for you know clear strong opinions for
[01:31:05] one or the other in our case Katherine
[01:31:08] it's an end and both for our existing
[01:31:13] demo technology as well as our our road
[01:31:16] map so we are working towards the 16t
[01:31:18] solution each Direction um uh sticking
[01:31:22] with eight fibers
[01:31:24] uh per Direction doubling the wavelength
[01:31:26] count to 16 upping upping the data rate
[01:31:30] and uh also looking at the modulation
[01:31:35] format okay so did I hear a number 16 T
[01:31:39] per Direction Ste tea okay thank you I'm
[01:31:42] I'm looking for some
[01:31:44] bidding um so Dan cooked I see you
[01:31:48] grinning do you would you like
[01:31:51] to you've you've been around the block a
[01:31:53] few times with this question so where
[01:31:55] where is your vote and what do you think
[01:31:58] I
[01:31:59] mean I
[01:32:01] I if if we if we looked at the cost of
[01:32:06] of optical interconnect and and cost
[01:32:10] where where the whole solution if we had
[01:32:14] the same data rate for the whole
[01:32:17] solution uh the multimode fiber pixel
[01:32:20] based stuff tends to come out cheaper
[01:32:23] but doesn't come out ahead in highest
[01:32:27] bits per
[01:32:29] fiber um so so I think you guys know
[01:32:33] where the number is it's it's like sub
[01:32:35] one terabit per fiber and and um but I
[01:32:39] still think that it has a tremendous
[01:32:43] opportunity um to be the lowest cost
[01:32:47] solution um in in that regard if if you
[01:32:50] you know you know pick a pick a solution
[01:32:54] I think we can find a a multim mode
[01:32:57] vixel based uh solution to it that's
[01:33:00] lowest cost interesting so just to be
[01:33:04] clear you know if your vo is for
[01:33:06] multifiber Solutions then how many how
[01:33:09] many vixel could you could you put
[01:33:11] together how many fibers would that be
[01:33:14] what's practical and then what does that
[01:33:16] get to you get to you to in terms of
[01:33:18] terabits I think traditionally it was 12
[01:33:20] in each direction for the vixel soltion
[01:33:23] yeah 12 I I think you know what what we
[01:33:25] were doing um 16 or
[01:33:29] 32 uh I already made I made the case for
[01:33:32] I think going Beyond 16 fibers in a
[01:33:35] connector or well Mt 24s are actually
[01:33:39] higher yield than than 16s right now um
[01:33:43] that that may be a historical thing um
[01:33:47] and I would be reluctant to go past 32 I
[01:33:51] I just think it's going to be um too
[01:33:54] costly I mean there's a whole aspect of
[01:33:57] high count fibers if fibers in general
[01:34:00] that I didn't bring up but I think you
[01:34:03] know Katherine as a former data center
[01:34:07] operator and and your your colleague
[01:34:09] Jeff Cox would say the connectors are
[01:34:13] the biggest problem and it's not even
[01:34:15] that they have low loss it's that people
[01:34:18] plug them in upside down sometimes and
[01:34:20] break them they get dirty um I I think
[01:34:24] you know you just don't want high fiber
[01:34:27] count in a connector you don't want it
[01:34:30] to they're harder to
[01:34:32] clean uh there's all sorts of reasons so
[01:34:34] I would sort of probably put a limit for
[01:34:36] both multi mode and single mode at 32
[01:34:40] although I don't like limits because
[01:34:41] somebody's going to come along and with
[01:34:43] a smaller connector or automate
[01:34:47] something I think as soon as it gets
[01:34:48] automated that changes probably our
[01:34:51] whole view on on these
[01:34:54] Channel counts but
[01:34:56] um at the time being in the near future
[01:34:59] that's where I would sort of put things
[01:35:01] yeah I you're raising some very good
[01:35:03] points there's some real world
[01:35:05] situations into you know to do with just
[01:35:08] running things operationally and
[01:35:10] infrastructure Rebecca you look like you
[01:35:12] had something you wanted to say there
[01:35:14] yeah so as someone who has spent eight
[01:35:16] years at Corning um I think I have a
[01:35:18] little bit of knowledge on fiber and
[01:35:20] fiber connectors and absolutely what Dan
[01:35:22] said you know there are challenges that
[01:35:24] hyperscalers see in the field in
[01:35:27] deployment with fiber connectors
[01:35:29] absolutely true um I do want to
[01:35:32] distinguish between the connectors that
[01:35:34] we're talking about though right when we
[01:35:37] go into AI systems and we're talking
[01:35:39] about co- package Optics type Solutions
[01:35:43] we are inherently talking about bringing
[01:35:45] fiber into the box and having
[01:35:47] connections that are only made during
[01:35:50] the creation of the
[01:35:52] box right and that's a different use
[01:35:54] case a different reliability in terms of
[01:35:56] the number of connections disconnections
[01:35:59] the type of technicians that are doing
[01:36:00] it whether automation can enable those
[01:36:03] connections to actually happen um that I
[01:36:06] think changes the dynamic within the box
[01:36:09] and changes the discussion within the
[01:36:11] Box than one that would be at the front
[01:36:13] panel so just something to think about
[01:36:16] there so you touched on something
[01:36:19] related to co-packaged Optics which I've
[01:36:21] been advocating for a while
[01:36:24] which is uh the the notion of field
[01:36:27] replaceability is a bit of a farce
[01:36:29] because the Assembly of the Box requires
[01:36:34] higher skill and um and even though
[01:36:37] things are put together
[01:36:39] once um it also requires higher skill to
[01:36:42] take them apart when they need replacing
[01:36:45] um and I think you're you're kind of
[01:36:47] alluding to that and I and I agree um I
[01:36:52] agree with that aspect
[01:36:54] um that's something we need
[01:36:56] to keep our eye on and I think Aid also
[01:37:00] said he feels that the connector is
[01:37:03] probably the least reliable thing uh in
[01:37:06] in the link so we have to keep our focus
[01:37:08] on that as well assuming to have
[01:37:11] reliable
[01:37:12] lasers I think I think this is a really
[01:37:15] a key point the application we
[01:37:17] addressing is really different than
[01:37:19] standard fiber optic applications right
[01:37:22] um we talking about short distances
[01:37:24] again where the fibers may be
[01:37:27] permanently attached we need much higher
[01:37:29] reliability because field replacement is
[01:37:31] impossible if it's co- packaged or board
[01:37:33] mounted as Dan pointed out but it really
[01:37:36] does allow you to use new things that
[01:37:38] you couldn't do before so we shouldn't
[01:37:41] be wedded to standard types of
[01:37:43] technologies that were really designed
[01:37:45] for other
[01:37:49] applications yeah I think that's
[01:37:51] certainly true in some of these new
[01:37:53] merging markets maybe in the front end
[01:37:56] um that there's there's enough single
[01:37:59] mode fiber out there we may end up with
[01:38:01] different solutions how interesting um
[01:38:04] anyone else got some more on this topic
[01:38:06] otherwise I have some some other themes
[01:38:09] uh from the questions um cost and power
[01:38:14] are coming up in as topics and um Rob in
[01:38:20] the chat has expressed this question a
[01:38:22] couple of different ways he but he's
[01:38:24] asking each of the panelists basically
[01:38:27] to pull holes in the others Solutions
[01:38:31] you know what do what do each of you see
[01:38:32] as the big disadvantage for doing it the
[01:38:35] other way right um and
[01:38:39] so I and then he's he's called out power
[01:38:43] and cost as two of the things that are
[01:38:46] top to his of his mind um I have a whole
[01:38:50] load of others but uh so Dan you you
[01:38:52] mentioned a little bit about cost yeah
[01:38:55] i' I'd like to go first I'd be more than
[01:38:57] happy to pick apart each of my fellow
[01:39:00] panelists uh Solutions once they hand me
[01:39:03] their bill of materials and and
[01:39:05] component costs because I have no idea
[01:39:08] what each of them cost right now um I
[01:39:12] have my I don't even know how to rank
[01:39:14] them in in order of
[01:39:19] cost okay you had to take an educated
[01:39:22] guest what it be right
[01:39:24] now what would
[01:39:26] it well well Allan you're selling lasers
[01:39:29] right so I'm going to give you you're
[01:39:31] the lowest cost just because you're a
[01:39:33] component vendor right um lasers
[01:39:36] interconnects we'll do both and and you
[01:39:40] know Peter and Amit are similar
[01:39:44] Technologies
[01:39:46] um I I I I probably want to put Peter
[01:39:50] next because he's a a small business
[01:39:52] that's R for customers and so he's going
[01:39:55] to be
[01:39:56] aggressive uh then Amit and then uh
[01:40:00] honestly Rebecca I wasn't clear
[01:40:02] on on what your solution was so I don't
[01:40:06] I don't know where to to put you I'm
[01:40:08] sorry don't don't all good all good and
[01:40:11] and by Design right we are new startup
[01:40:15] um there's a lot I cannot uh say
[01:40:18] publicly as to what we're doing um but
[01:40:21] you know maybe the answer the question a
[01:40:23] little bit more generically um one of
[01:40:26] the things that I would say you're
[01:40:28] looking at when you're driving to AI is
[01:40:30] also an opportunity cost so if you can't
[01:40:32] scale to the volume that AI can scale to
[01:40:36] then your solution is costing the
[01:40:39] customer more in the end even if your
[01:40:42] individual component costs you know less
[01:40:45] than a cent if you can't scale to the
[01:40:47] volume you've missed out on a huge uh
[01:40:51] Revenue opportunity and customer cannot
[01:40:54] deploy which is a huge cost to the
[01:40:58] customer so if they're going to bet on
[01:41:00] you they need to bet on something that
[01:41:01] can scale to that volume which I think
[01:41:03] is really important one of the
[01:41:06] challenges I see with more wavelengths
[01:41:08] is there is not a proliferation of
[01:41:11] sources that can be used for high power
[01:41:14] to enable the solution and uh no offense
[01:41:19] to our fellow speaker Allan but pretty
[01:41:21] much all of the multi-wavelength sources
[01:41:24] aside from AIT are
[01:41:28] startups and you know if you look at
[01:41:31] that as a solution space to work within
[01:41:35] then it becomes very hard to bet on a
[01:41:37] solution and know that it will scale
[01:41:40] especially when it's so material
[01:41:42] dependent and history tells us that
[01:41:44] materials take a long time to get to
[01:41:46] full reliability and volume requirements
[01:41:49] so that's where I would see a large
[01:41:51] challenge uh in scaling and uh as much
[01:41:55] as AIT loves his Laser Technology it's
[01:41:57] not available to
[01:41:59] everyone I I think Rebecca just invited
[01:42:02] me to uh segue into your question about
[01:42:04] manufacturing ramping Catherine but um
[01:42:06] quick comment so on the topic of high
[01:42:09] power I think that's actually the wrong
[01:42:10] direction to go for lasers because laser
[01:42:12] reliability actually scales one over
[01:42:14] current to the fourth power or third so
[01:42:18] going using high powered lasers and
[01:42:20] splitting isn't the optimal direction
[01:42:22] for syst level reliability and I think
[01:42:25] um the general direction that the
[01:42:28] industry is going in terms of you know
[01:42:30] using it more and more power higher
[01:42:32] power lasers for dr4 dr8 splitting um
[01:42:37] isn't necessarily the best approach in
[01:42:39] my mind because again of that
[01:42:41] reliability you're you're literally
[01:42:43] going the opposite direction of where
[01:42:45] lasers are more reliable which is lower
[01:42:47] power and lower currents um and also if
[01:42:49] you knock out that one laser goes down
[01:42:51] all all four or eight l go down so um
[01:42:55] the blast radius is quite high and so
[01:42:57] the use case for us with comb Blazers
[01:42:59] isn't to um try
[01:43:03] to uh put out a high very high power
[01:43:06] from the laser itself because of that
[01:43:08] reliability consideration but there's
[01:43:09] other ways you can compensate for that
[01:43:11] in the link itself and then on the topic
[01:43:13] of manufacturing ability um you know I
[01:43:15] think Rebecca brought up great points
[01:43:17] it's um you know laser material and
[01:43:20] reliability takes time we've been
[01:43:22] working working on that um either
[01:43:25] directly at contant or indirectly at
[01:43:27] UCSB for the better part of a decade now
[01:43:29] and so we feel pretty good about
[01:43:30] material supply chain we have multiple
[01:43:32] partners for the Quan EP um that we
[01:43:36] believe we can go to volume with and and
[01:43:38] we understand the physics of the
[01:43:40] reliability and failure me mechanisms
[01:43:43] pretty well so um at least for quintant
[01:43:46] um we feel pretty good about you know
[01:43:47] the concerns that were brought
[01:43:50] up yeah I think I like uh want to build
[01:43:53] on Alan's Point U traditionally I mean
[01:43:58] there's two been two reasons to go to
[01:44:00] very high powered lasers one I think Bia
[01:44:03] mentioned is the threshold penalty you
[01:44:05] don't get it to L so when you go to a
[01:44:08] shared laser you pay the threshold
[01:44:09] penalty once for as many channels as
[01:44:12] you're driving the second is you know
[01:44:14] the cost of a laser and everything that
[01:44:16] goes into it um and typically the power
[01:44:19] is not scaling with area but the cost is
[01:44:21] scaling with area so so I think that
[01:44:24] that's been the traditional reason but
[01:44:25] in some of the new approaches with QD
[01:44:27] lasers in our particular case we have a
[01:44:30] certain size um Indian fosite coupon we
[01:44:33] can get on there and so you know multi
[01:44:36] Lambda for us comes pretty cheap it's
[01:44:38] not equal to the cost of adding a
[01:44:40] separate laser um and uh you know in
[01:44:43] terms of the power and how you address
[01:44:45] the link budgets another thing that we
[01:44:49] like to use uh selectively and
[01:44:53] inventively is s soas right you don't
[01:44:56] need all the power coming out of your
[01:44:57] source that you need to close the link
[01:45:00] so so you know how you use S soas in the
[01:45:02] mix also makes the overall solution
[01:45:04] architecture very
[01:45:11] interesting yeah so this is it's a great
[01:45:14] discussion I think we could go on doing
[01:45:15] this all evening but we do it looks like
[01:45:18] the poll has converged so um I think we
[01:45:21] have a result and um I was expecting it
[01:45:26] to be much closer than this it looks
[01:45:28] like we actually have
[01:45:30] 43% of the voters voted for
[01:45:34] wavelength Alan that's despite you being
[01:45:36] hugely outnumbered I feel um but the
[01:45:40] underdog
[01:45:42] wins um 19% voted for fiber but uh there
[01:45:49] is a very large contingent that's still
[01:45:51] undecided
[01:45:53] so whether that's uh all those other
[01:45:56] catch things in there whether it's both
[01:45:58] or maybe or it depends or one first then
[01:46:01] the other who knows I think there's
[01:46:03] still plenty to discuss here um so thank
[01:46:07] you very much I'm going to stop sharing
[01:46:09] the poll now and I'm looking at Dan I'm
[01:46:11] wondering how much of this sort of Grace
[01:46:13] extra time do you think we can really
[01:46:15] borrow here
[01:46:17] Dan can't hear
[01:46:20] you as long as the Ence doesn't
[01:46:22] disappear you can go a little longer
[01:46:25] should we do some more of this I'd say
[01:46:27] let's do one more question yeah one more
[01:46:29] question okay um I get to pick because I
[01:46:33] don't see any more questions here but um
[01:46:36] how how could you possibly we were
[01:46:38] talking about radic a lot so this is
[01:46:41] this is an an odd question I'm going to
[01:46:44] ask Rebecca since you talked a lot about
[01:46:48] radex how would this work with Optical
[01:46:51] switching
[01:46:53] right it so um we have Optical switching
[01:46:57] hasn't happened except Google right and
[01:46:59] Google's special in a lot of ways but I
[01:47:02] hear Optical switching being talked
[01:47:04] about um there's a limited number of
[01:47:08] mems mirrors that can move data around
[01:47:13] in that switch it's a bit of a special
[01:47:15] case but how would you do that if you
[01:47:17] had lots of parallel fiber um or lots of
[01:47:22] different fi is going around how would
[01:47:24] that work um maybe to answer this in a
[01:47:27] different way just because I can't
[01:47:29] divulge some of the stuff we are working
[01:47:31] on um but what not to do right so one of
[01:47:35] the biggest challenges with Optical
[01:47:39] circuit switching is the
[01:47:43] loss so if you take loss into account
[01:47:47] you say you now you know let's say you
[01:47:49] had a Dr Link you're now penalizing
[01:47:52] yourself in that Dr link right you're
[01:47:55] taking away some of the connections that
[01:47:57] you would have normally been able to do
[01:47:59] in order to support the solution so you
[01:48:03] have to be very focused then when you
[01:48:06] deploy the system of meeting the link
[01:48:08] budget required which makes that sort of
[01:48:11] optical circuit switching very very
[01:48:13] challenging right I think a Google
[01:48:16] solution is something like 3db uh link
[01:48:19] loss um and a Dr Link budget's 4db
[01:48:23] right so they're pushing the transceiver
[01:48:25] vendors on either end to have a slightly
[01:48:27] larger link budget than the drpc in
[01:48:30] order to compensate um they're also um
[01:48:33] making both directions go on the same
[01:48:35] fiber and having to have recirculators
[01:48:38] to pull out each Direction and separate
[01:48:40] them which adds to back reflection
[01:48:42] tolerance challenges all to get around
[01:48:45] the fact that you can't put so many
[01:48:47] little mirrors in a place at a given
[01:48:49] time right so um definitely I think a
[01:48:53] challenge to use that technology and
[01:48:56] scale if you take Optical circuit
[01:48:58] switching on Silicon
[01:49:01] specifically there's a lot of challenge
[01:49:03] there because you want to use single
[01:49:05] mode
[01:49:07] technology um and single mode technology
[01:49:09] means that you have both uh
[01:49:12] polarizations entering your silicon and
[01:49:15] now you have to figure out how to do
[01:49:16] circuit switching with both
[01:49:18] polarizations which adds to loss because
[01:49:20] you need to either figure out how to put
[01:49:22] those polarization ation back to a
[01:49:23] single polarization without uh causing
[01:49:26] noise to your signal um or you have to
[01:49:29] have two separate completely separate
[01:49:30] circuits operating independently which
[01:49:32] adds to the size of the device and the
[01:49:34] power of the device which doesn't make
[01:49:36] any sense so right there you're probably
[01:49:39] at a loss that's even higher than the
[01:49:41] 3db loss that I just mentioned Google
[01:49:44] has because not only do you have to deal
[01:49:45] with the polarization rotation but every
[01:49:47] single one of those little itsy bitsy
[01:49:49] circuit switches that you want to put in
[01:49:51] there to make a non-blocking s solution
[01:49:53] is adding 0.1 db2 DB and so forth that
[01:49:57] adds up dramatically into that solution
[01:50:00] um really heavily hitting onto the
[01:50:02] overall link budget that you have and
[01:50:04] the power that you would need in a laser
[01:50:06] uh to compensate for it so I definitely
[01:50:09] think those are very challenging um
[01:50:11] solutions that you need to address in a
[01:50:14] completely different way that
[01:50:15] unfortunately I can't discuss here um
[01:50:18] but you know the these are compounded um
[01:50:22] if you also add wavelength into the mix
[01:50:24] right um even further if you do those in
[01:50:27] a traditional Optical C circuit
[01:50:29] switching method there there is one one
[01:50:32] other circuit switch that you didn't
[01:50:33] mention Rebecca which is the simple uh I
[01:50:36] call it low Tech but efficient Tech the
[01:50:39] uh the the fiber robot that just sure um
[01:50:43] I considered mentioning that which is
[01:50:45] actually not that bad because what what
[01:50:47] what the optical circuit switch is is
[01:50:49] really a an automated patch pan it's not
[01:50:51] like a c switch in the in the
[01:50:55] sonit it's an automated patch panel and
[01:50:59] for that speed doesn't matter all that
[01:51:01] much and uh once you start about uh
[01:51:03] talking about connecting fibers you can
[01:51:06] connect empty feres that's fine then you
[01:51:08] have with the exact same complexity and
[01:51:10] loss you all of a sudden connected 32
[01:51:12] fibers to each
[01:51:14] other uh yeah so I I would say at that
[01:51:16] point right you're talking about uh
[01:51:18] space in the data center right because
[01:51:20] it is relatively large right it's a kind
[01:51:23] of a full rack type solution um and
[01:51:27] whether or not you can fit it in the
[01:51:29] data center footprint um or you'd rather
[01:51:31] use that for GPU as their better
[01:51:33] solution you can get even denser on but
[01:51:35] yeah I would agree with you that it adds
[01:51:37] just a connection Point yeah in that
[01:51:40] sense well it was kind of an
[01:51:42] off-the-wall question I wasn't expecting
[01:51:44] any of you to have prepared for that one
[01:51:46] so uh thank you very much for being such
[01:51:49] huge Sports this was a lot of fun I
[01:51:52] would like to thank each of you for
[01:51:54] joining this has been wonderful and
[01:51:56] thank you for really getting into it and
[01:51:58] and you know sharing all all the details
[01:52:01] and the nitty-gritty of why why this is
[01:52:05] or isn't the right way to do it and um I
[01:52:09] think um let's see
[01:52:12] uh there's not much else to say except
[01:52:15] thank you very much I think you you
[01:52:17] moved a lot of the audience which is
[01:52:19] very interesting and um man over to you
[01:52:23] I think there's so one more one more
[01:52:26] final thing to say right there is first
[01:52:28] of all I want to thank you Katherine for
[01:52:30] moderate agreeing to moderate this
[01:52:32] session doing a spectacular job pulling
[01:52:33] it together and running it this is the
[01:52:36] best panel I think we've had in 20 years
[01:52:38] 25 and uh I just thrilled with the the
[01:52:44] inventive people we have on the panel
[01:52:46] and know every one of you has has
[01:52:49] brought really new thinking into my
[01:52:51] world and of course you know I've known
[01:52:53] most of you for for a while and I am
[01:52:55] excited by the future of optical
[01:52:58] technology especially in the data center
[01:53:00] and to accelerate AI um I didn't vote
[01:53:04] because I'm not allowed to um and and I
[01:53:08] I kind of liked what Allan said which is
[01:53:10] you've got the fibers there let's you
[01:53:12] know add to the wavelengths first but I
[01:53:14] also think with some of the new uh fiber
[01:53:17] multiplexing um and parallel
[01:53:19] Technologies coming out and especially
[01:53:21] perhaps with you know lowcost LEDs we're
[01:53:23] going to see more opportunities to do
[01:53:25] that in in parallel and the data center
[01:53:27] challenges are so different from the
[01:53:29] Long Haul Telecom challenges but I think
[01:53:32] it's right for for some new kinds of
[01:53:33] approaches to using using Optics so uh
[01:53:37] my final word is uh well first of all we
[01:53:40] have uh another great day starting
[01:53:42] tomorrow at nine o'clock but um H
[01:53:45] interconnect started in I think 1993
[01:53:47] I've been involved since a year or two
[01:53:48] after that and we always met uh in
[01:53:51] person on the Stanford campus until 2000
[01:53:54] through
[01:53:55] 2008 um and then we uh we moved to New
[01:53:59] York for one year and then we came back
[01:54:01] to the Bay Area and instead of being at
[01:54:03] Stanford we were at companies uh the
[01:54:05] reason we went to New York and by the
[01:54:08] way well this is hard to tell this in in
[01:54:10] the right sequence um we used to have
[01:54:13] sometimes dinner before the panel
[01:54:14] discussion or the cocktail reception
[01:54:16] before the panel discussion which case
[01:54:17] the panel got really interesting um but
[01:54:20] we always concluded the first day with
[01:54:21] some kind of um a cocktail uh period and
[01:54:25] a cocktail hour so in 2009 we went to
[01:54:29] New York and we're housed by hosted by
[01:54:32] credit
[01:54:33] swis and the reason for that is because
[01:54:35] we had uh a couple of attendees from
[01:54:37] credit s and one from the New York Stock
[01:54:39] Exchange that would come to our
[01:54:41] conference every year at Stanford um and
[01:54:44] one in particular sat in the back and
[01:54:46] didn't say too much but whenever he
[01:54:48] spoke it was like oh boy we didn't think
[01:54:50] of that his name was head Bubba or the
[01:54:53] head Bubba that was his legal name it
[01:54:56] was not the name his parents gave him at
[01:54:58] Birth but it became his legal name and I
[01:55:01] actually have a photograph of
[01:55:10] him not a great one he didn't look like
[01:55:14] we normally expect the V Vice President
[01:55:16] of it at a at a major Swiss bank credit
[01:55:19] swis to look but that was him
[01:55:23] um and he uh he brought a lot of reality
[01:55:28] to our discussions and said you know I'm
[01:55:30] running a bank here this is what's
[01:55:32] important to me and you're talking in
[01:55:34] theory or whatever wasn't relevant um
[01:55:37] and so we listened to him and then he
[01:55:39] hosted us and we were very grateful and
[01:55:40] then we came back here um and he was a
[01:55:45] very generous guy uh he shared things
[01:55:48] with us he had certain hobbies and
[01:55:50] interests but he was very generous and
[01:55:52] and very kind he died in 2013 of a
[01:55:54] medical problem and since then we have
[01:55:57] been memorializing him at the cocktail
[01:56:00] reception right after or before the
[01:56:03] panel well since 2020 we haven't been in
[01:56:06] person but we don't want to let this
[01:56:08] opportunity go to waste to thank him for
[01:56:10] his friendship and to remember him and
[01:56:11] so here I have happen to have handy I
[01:56:15] should probably turn off the uh virtual
[01:56:17] background a fine glass of California
[01:56:21] Zinfandel I'd like to make a toast to
[01:56:23] the memory of the contributions and the
[01:56:25] kindness of the head
[01:56:32] Baba and with
[01:56:34] that we shall end today's program look
[01:56:38] forward to a great one tomorrow and
[01:56:39] thank you all for the best panel in just
[01:56:42] about living
[01:56:45] memory thank you everyone thank you Dan
[01:56:47] and
[01:56:48] Katherine and thanks all to the audience
[01:56:50] who stuck around for a long time
[01:56:53] bye-bye bye