Full Transcript
https://www.youtube.com/watch?v=4UCHG5EmgH4
[00:01] Okay.
[00:01] Okay.
[00:03] Oh, thank you very much for the kind introduction.
[00:05] So, the my name is Hideyuki Nasu from Furukawa Electric, Japan.
[00:10] So, I'm very happy to be here here in in Barcelona.
[00:12] Really good, you know, landscape and you know, good, you know, air.
[00:17] So,
[00:19] So, the title of my presentation is the energy efficient VCSEL-based transceiver for AI data centers.
[00:26] So, I should use this, okay.
[00:29] So, this illustration is showing that you know, schematic of the AI you know, data center network.
[00:37] So, actually we have the legacy you know, data transmission network.
[00:41] We are saying that this is front-end network now.
[00:44] And for the computing side, we have a back-end network.
[00:46] Here we have a scale-up you know, network and a scale-out inter-building and a scale-out inter-building.
[00:52] For the long-distance trans you know, the network, we're using the you know, optics, with particularly
[01:01] the you know, optics, with particularly pluggable optics and for the multiple.
[01:03] pluggable optics and for the multiple port and high-density optical.
[01:05] port and high-density optical interconnect,
[01:07] interconnect, uh the adoption of the CPO has been expected like this.
[01:12] expected like this. So, then the next target for the optics should be the this scale-up.
[01:17] interconnect, but you know, of course it is very high speed.
[01:21] So, today we have this kind of you know, scale-up network.
[01:24] Right now, we are using the copper cables for the GPU-GPU connection or GPU to your network switch.
[01:29] and the GPU. So, actually at the you know, exhibition, I saw the Amphenol's interesting the high-density electrical wiring system.
[01:42] and they have a very, I mean, small diameter. It looks like you know, fiber cable.
[01:47] So, then you know, if the distance is something like you know, 2 m, it is okay to use a copper cable.
[01:55] Then you know, data rate is coming to the 200 gig per lane now, and power consumption will be 5 to 10 picojoule.
[02:03] consumption will be 5 to 10 picojoule per bit including the solid state devices.
[02:08] Then the future, you know, our, you know, scale-up system have to be in the wider, then the distance of the, you know, interconnection should be the longer.
[02:15] So then the target will be, you know, speed is a consistent like a 200 gig, maybe or more than 200, you like a 400 gig will be required, and the reach will be around 20 m, and the power should be the consistent.
[02:28] We cannot increase the power.
[02:30] So then, you know, we expect in the adapt to the optics to the scale-up network.
[02:35] Then what is the candidate?
[02:37] Then now in the people are talking about the big cell based technology for the scale-up because big cell is providing a very low power.
[02:45] So then there is some demonstration of the 200 gig big cell, and you can take a look at this, you know, this is the reporting from, you know, Broadcom.
[02:54] They're showing the modulation rate as a function of the year.
[02:59] So then at the 2022-2023, they, you know, reach at the 100 gig per
[03:04] they, you know, reach at the 100 gig per lane.
[03:07] So now we are 2026 is the time to demonstrate, you know, 200 gig.
[03:10] So actually they showed you this, you know, 100 gigabit PAM4 eye diagram.
[03:15] So then it looks, you know, 200 gig is a feasible.
[03:18] And also there's approach to use the 1060 nm.
[03:23] There is, you know, some discussion about the standardization of the wavelengths of 1060 nm for 200 gig optical link.
[03:34] And this demonstration is particularly 1060 nm single mode big cell using the coupled cavity.
[03:40] Coupled cavity is good to expand the modulation, you know, bandwidth.
[03:45] And then they demonstrated the 12 channel multi by 4 to the matrix, then, you know, they achieve the uh 200 gig uh you know, PAM4 signal for all 48 channels.
[04:02] So, then just like to revisit uh
[04:05] So, then just like to revisit uh short-line density as a kind of you.
[04:08] short-line density as a kind of you know, use case um.
[04:10] know, use case um like uh NVLink 5 as a NVLink 72, then uh.
[04:15] like uh NVLink 5 as a NVLink 72, then uh the bandwidth should be 14.4 terabits per second.
[04:22] So, then we have a 100 mm uh millimeter for the short line.
[04:25] So, then uh the density should be uh 0.14 terabits per second per millimeter.
[04:31] So, then if we put the uh 3.2 terabits per second transceiver of eight pieces uh um in the edge of the uh uh this you know, GPU, that would be you know, capable.
[04:43] So, then uh actually we have a particular uh you know, transceiver and we see a you know, very small uh size I will show the later and uh that would be capable to to uh allocate to this kind of you know, structure.
[05:00] So, then actually uh for the you know, electrical interface, we demonstrated integrated uh electrical.
[05:06] demonstrated integrated uh electrical pluggable interface.
[05:09] This one you know, pluggable interface.
[05:09] This one you know, the interface accommodate eight you.
[05:11] the interface accommodate eight you know, transceivers.
[05:14] Then the know, transceivers.
[05:14] Then the uh the length is 92 mm.
[05:18] It's capable to uh the length is 92 mm.
[05:18] It's capable to adapt to 100 m you know, edge.
[05:21] And also we demonstrated CPO daughterboard using.
[05:23] we demonstrated CPO daughterboard using the full.
[05:25] the full uh you know, electrical pluggable.
[05:26] uh you know, electrical pluggable interface and uh this can accommodate.
[05:29] interface and uh this can accommodate the 32.
[05:30] the 32 uh CPO transceivers.
[05:34] So, um let's think about uh NVLink 6 and.
[05:38] So, um let's think about uh NVLink 6 and every L 40 uh 144.
[05:43] So, then that you see a 28 terabits per second board.
[05:46] So, then, uh.
[05:48] then, uh density will be 0.28 terabit per second.
[05:51] per millimeter.
[05:52] So, then,
[05:54] So, then, the idea to to, you know, increase the.
[05:57] the idea to to, you know, increase the boundaries, we can allocate the.
[06:00] boundaries, we can allocate the transceiver like, you know, eight pieces.
[06:02] transceiver like, you know, eight pieces in line.
[06:02] So, then, we, you know, use the this two rows.
[06:05] So, then, eight pieces
[06:08] This two rows.
[06:08] So, then, eight pieces multiply two rows.
[06:10] So, then, we can multiply two rows.
[06:12] So, then, we can achieve the achieve the uh.
[06:13] Uh let's say, accommodate 16 in transceiver.
[06:16] Let's say, accommodate 16 in transceiver of the one edge.
[06:19] So, then, the density will be 0.51 terabit per second per millimeter.
[06:22] Millimeter.
[06:23] So, um.
[06:25] So, um let's see.
[06:26] Let's see.
[06:26] To use the big cell, some people have a concern about the reliability.
[06:31] But, the actually, like I hold the silicon photonics transceiver, the people are using external laser source.
[06:35] You know, external laser source is located in the pretty good, you know, environmental condition where, you know, temperature is pretty low.
[06:46] Actually, it it is here attached to the, you know, front panel.
[06:49] So, then, but um I'm working for the, you know, laser devices long time.
[06:55] And you know, temperature is a key.
[06:57] And also, the, you know, the current density is very, very important.
[07:02] So, then, what do we do?
[07:04] Um you know, this is the uh.
[07:06] We have the this, you know, the cooling.
[07:09] We have the this, you know, the cooling system.
[07:13] The particularly, we have the system.
[07:13] The particularly, we have the you know, switch ASIC.
[07:16] The data have you know, switch ASIC.
[07:16] The data have you know, this a cooling with the cold plate.
[07:18] know, this a cooling with the cold plate there on the top.
[07:20] Then, you know, transceiver is here kind of the bottom.
[07:22] So, we attach the, you know, cold plate.
[07:25] the bottom side of the this, you know, daughterboard.
[07:29] So, we could simulate, you know, temperature distribution, but.
[07:33] we tried to make the, you know, experimental setup like this.
[07:38] And to to, you know, imitate.
[07:40] uh.
[07:41] uh uh let's say, the switch.
[07:43] we use the uh, you know, ceramic heaters.
[07:45] and also the uh, you know, transceiver is used the uh, you know, ceramic trans-.
[07:51] uh,
[07:52] ceramic uh, heaters.
[07:54] So, then what we do is uh.
[07:58] So, actually this, you know, graph showing the uh, uh, operating temperature of the switch.
[08:02] ASIC as a function of the power consumption of the switch ASIC.
[08:06] If we, you know, increase the power for the switch ASIC, it is obviously the
[08:10] switch ASIC, it is obviously the operating temperature is increased.
[08:12] operating temperature is increased.
[08:15] However, the transceiver, that case temperature of the transceiver is very much consistent.
[08:19] much consistent.
[08:21] So, then uh, we can achieve like a big cell temperature is around like uh less than 45.
[08:26] less than 45.
[08:29] So, just thinking about the reliability for the big cell, in 40-50 this range, the single to big cell has a 0.03 0.3 a 0.05 level.
[08:37] Then, if we have the 30 channel, it is around the one.
[08:42] So, then, you know, we can keep the good reliability of the big cell with this cooling, you know, system.
[08:46] this cooling, you know, system.
[08:49] So, let's think about the link energy of big cell based CPU transceivers.
[08:55] So, then uh, there is seeing graph uh, link energy expressed with a p- a picojoule per bit as a function of the bit rate.
[09:00] per bit as a function of the bit rate.
[09:03] There's a approach uh, to to reduce the power consumption.
[09:06] But, actually, big cell itself is a very low bias current and uh, you know, the threshold current
[09:11] and uh, you know, the threshold current is be very um, the small.
[09:17] So, then, what is the uh, you know, cause of the power consumption?
[09:21] It is basically driver field devices, big cell drivers and the TIA.
[09:24] and there's some, you know, report um, the demonstration and using the you know, BiCMOS or, you know, CMOS.
[09:32] Then, I tried to plot, you know, data uh, regarding the picojoule per bit.
[09:37] So, then CMOS device is here, you know, the good, actually.
[09:40] I mean, that uh uh there is a power consumption and there is uh you know, the significant demonstration of this.
[09:48] This is 0.9 picojoule per bit.
[09:52] Uh this is demonstrated by Intel.
[09:54] They're using the uh you know, the CMOS.
[09:58] And then actually, they have uh three different, you know, this voltage.
[10:03] Mainly, they use 0.9 V.
[10:04] So, kind of special, but uh you know, demon- demonstrated you know, this kind of you know, very small uh uh picojoule per bit.
[10:13] uh picojoule per bit.
[10:16] So, let me try to uh com- uh make a video comparison of the uh you know, the performance and the cost regarding the copper cable and the BICSI and silicon photonics technologies.
[10:24] photonics technologies.
[10:26] So, then uh what do you Here is the uh you know, copper including DAC, ACC, and AEC, BICSI, and the silicon photonics.
[10:35] So, which obviously, you know, that the copper cable is short.
[10:37] BICSI is capable with a 30-m, you know, with a multi-mode fiber.
[10:43] But if we move to the single mode, that will be extended to the 500-m even it is 200-gig.
[10:50] So, um 200-gig operations, those are you know, doesn't matter.
[10:53] So, then the link energy including SerDes.
[10:55] So, the BICSI is uh you know, only like a 0.9 picojoule per bit.
[11:02] Then including the SerDes, it's still like a 5 picojoule per bit.
[11:07] It's quite low.
[11:09] So, then um but reliability is a bit concerned, but actually, I am as mentioned, you know,
[11:15] actually, I am as mentioned, you know, temperature control is very important.
[11:17] temperature control is very important thing to do to realize a good thing to do to realize a good reliability.
[11:23] So, um let me introduce the our approaches.
[11:27] you know, introduce the our approaches.
[11:29] Actually, we worked for the particular national project in Japan.
[11:32] particular national project in Japan.
[11:35] And this one was uh you know pretty high dense you know optical transceiver.
[11:37] dense you know optical transceiver.
[11:40] Actually we use the 16 channel pixel array and the photo diode array and the back back illuminated and bottom emitting you know pixel.
[11:42] array and the photo diode array and the back back illuminated and bottom emitting you know pixel.
[11:44] back back illuminated and bottom emitting you know pixel.
[11:47] Then those are pretty bonded interposer and we have the you know about the coupling with the multi-core fiber.
[11:51] pretty bonded interposer and we have the you know about the coupling with the multi-core fiber.
[11:52] you know about the coupling with the multi-core fiber.
[11:55] Actually we use the 19 core multi-core fiber.
[11:58] core multi-core fiber.
[12:01] The core is controlled let's say designed for the 1060 nanometer.
[12:02] 1060 nanometer.
[12:05] And you know this is anyway the butt joint and the bottom side of the interposer we have pixel driver and the TIA.
[12:07] joint and the bottom side of the interposer we have pixel driver and the TIA.
[12:10] interposer we have pixel driver and the TIA.
[12:12] Also these devices are the flip chip bonded on the interposer.
[12:14] devices are the flip chip bonded on the interposer.
[12:14] And we using the 0.3 mm land
[12:18] Interposer.
[12:21] And we using the 0.3 mm land grid array and also to realize the grid array and also to realize the pluggable transceiver.
[12:23] Pluggable transceiver I mean pluggable electrical interface we use the you know contact probes.
[12:28] For the you know realize pluggable.
[12:30] You know realize pluggable.
[12:33] So then what to bring it is 0.3 mm by 15.9 mm.
[12:37] It's kind of special right because we are using the multi-core fiber and using LC connector for each.
[12:43] Multi-core fiber.
[12:45] So then we also have the approach to use the existing you know product.
[12:48] So using the HCL but you know number of channels are reduced but it is 1060 nanometer top emitting 250 micron pitch in the pixel array can be used.
[12:59] Used.
[13:00] And we can use a single mode you know empty ferules.
[13:07] So then this is the existing interface like a for the QSFP-DD or something and also we have you know 1060 nanometer single mode fiber for sensing usage or.
[13:17] Fiber laser and also you know we
[13:20] fiber laser and also you know we Furukawa is manufacturing 980 nanometer fibers for the pump lasers.
[13:26] fibers for the pump lasers.
[13:29] mode fiber it doesn't matter.
[13:32] um let's see.
[13:33] Number of channel is the you know eight channel, but the this should be kind of um
[13:36] um user-friendly approach, I think.
[13:39] user-friendly approach, I think.
[13:41] So, let me show the some you know data.
[13:44] Uh this is 16 channel uh uh the transmission.
[13:47] the transmission.
[13:49] You can see the uh optical eye diagram of the 16 channel and uh electric eye diagram of the channel.
[13:53] The modulation speed is the 50 gig uh and everything.
[13:59] And also uh this is the you know bubble sub curve.
[14:01] We can achieve the less than 2.4 by 2 uh 2 2 minus 4.
[14:07] So, that should be the less than the KP4 FEC threshold.
[14:10] And also we demonstrated this in one 106.25 gig with a PAM4.
[14:15] And uh we transmitted the uh
[14:20] the uh multi-core fiber of the 2 km.
[14:23] multi-core fiber of the 2 km.
[14:23] Then we have the this unit optical eye diagram.
[14:25] have the this unit optical eye diagram.
[14:25] And then we achieve the uh less than a bit error rate of the less than KP4 FEC threshold.
[14:27] And then we achieve the uh less than a bit error rate of the less than KP4 FEC.
[14:29] bit error rate of the less than KP4 FEC threshold.
[14:29] Then we achieve the 3.95 pJ per bit.
[14:33] threshold.
[14:33] Then we achieve the 3.95 pJ per bit.
[14:34] This thing, you know, transceiver just have a single DC voltage, which is the 3.3 V.
[14:36] transceiver just have a single DC voltage, which is the 3.3 V.
[14:36] Also, this transceiver is including microprocessor.
[14:40] voltage, which is the 3.3 V.
[14:40] Also, this transceiver is including microprocessor.
[14:41] Also, this transceiver is including microprocessor.
[14:43] microprocessor.
[14:43] They including all you know controls in and they using the uh look up table.
[14:46] They including all you know controls in and they using the uh look up table.
[14:48] and they using the uh look up table.
[14:48] So, we achieved the this you know pJ per bit.
[14:50] we achieved the this you know pJ per bit.
[14:51] bit.
[14:51] So, then another approach is uh you know eight channel.
[14:53] So, then another approach is uh you know eight channel.
[14:53] Actually, we use a 980 nm single mode fiber, then transmitted the uh you know the fiber with the different lengths like a 500 m, 1 km, 2 km.
[14:56] Actually, we use a 980 nm single mode fiber, then transmitted the.
[14:59] single mode fiber, then transmitted the uh you know the fiber with the different lengths like a 500 m, 1 km, 2 km.
[15:01] uh you know the fiber with the different lengths like a 500 m, 1 km, 2 km.
[15:01] So, then even the transmitting 2 km single mode fiber, the the Q value is the 2.53.
[15:06] lengths like a 500 m, 1 km, 2 km.
[15:06] So, then even the transmitting 2 km single mode fiber, the the Q value is the 2.53.
[15:08] then even the transmitting 2 km single mode fiber, the the Q value is the 2.53.
[15:11] mode fiber, the the Q value is the 2.53.
[15:11] And uh you know bit error rate 8 is a 4 5.4 by 10 to minus 6.
[15:14] And uh you know bit error rate 8 is a 4 5.4 by 10 to minus 6.
[15:19] 5.4 by 10 to minus 6.
[15:19] Then evaluation of optical link uh let's
[15:22] Then evaluation of optical link uh let's say this is 53 gig ball.
[15:24] Say this is 53 gig ball.
[15:28] And this is the 106 gig.
[15:31] So then anyway, we can transmit the signal in the, you know, 2 km single mode.
[15:37] And the link energy for the 106 gig is as small as 4.1 picojoule per bit.
[15:42] So 4 picojoule per bit it's the you know, capable at least the the kind of fully packaged the, you know, optical transceiver.
[15:53] So then let's say but uh you know, the 90 nanometer single mode fiber is existing.
[16:01] But as I mentioned, you know, 1060 nanometer multi-core fiber is kind of customized and very special at this moment.
[16:09] So maybe we need to the discussion about the standardization of the fiber.
[16:14] And also the we need to demonstrate the 200 gig.
[16:17] Actually, I showed the 100 gig, you know, data, but uh 200 gig we are working on it.
[16:23] working on it.
[16:25] So then we need to the the demonstrate the 200 gig per lane optical link.
[16:29] uh optical link.
[16:32] And yeah, hopefully, you know, we can be commercialized with this in a big cell product.
[16:36] But the whole two adopt to the scale-up network, manufacturability is very important.
[16:39] It It must be huge in mass production.
[16:44] So that will be very very important.
[16:45] very important.
[16:45] Okay, that's it.
[16:45] Thank you very much.
[16:51] [applause]
[16:51] [applause]
[16:54] Are there any questions?
[16:54] Please.
[17:01] Okay, thank you very much for your presentation.
[17:04] I think single mode so it's a very promising technique.
[17:06] But can I ask what is your target use case of this?
[17:12] If you consider about using single mode in the scale-up network, actually we don't need such a long distance, right?
[17:17] Mhm.
[17:19] But, if we use it in a scale out network, that 100G and 200G standard has already been almost
[17:25] 200G standard has already been almost published.
[17:25] Published.
[17:28] Mhm. So, at least in actually, it seems that there's no nowhere to get it standardized.
[17:33] So, um what is your strategy to make it more promising to the to the industry?
[17:40] Yeah, thank you very much for your question.
[17:44] Actually, the firstly, we are looking at the scale up network, but I you know, data rate should be 200 gig.
[17:49] So, then um basically, we have a requirement about the realization of the low power.
[17:55] So, then you know, big cell based technology should be very good.
[17:58] Then why we are doing the single mode?
[18:02] It is like to I mean, uh realize full, you know, data center.
[18:06] If we have the a 100 gig per line, so we can achieve 2 km.
[18:11] You know, 2 km is a general requirement of the full frame, you know, data center, inter data center.
[18:15] So, then um we are also looking at, I mean, the scale out network.
[18:17] So, then let's see, you know, as I mentioned, you know, compared to with the silicon photonics transceiver, the big cell
[18:25] photonics transceiver, the big cell based transceiver can realize a very low power.
[18:31] So, then we like to, I mean, uh let's say the explain this.
[18:34] So, then elaborate, you know, the benefit of the big cell to the, you know, customers.
[18:39] So.
[18:40] Okay, thank you.
[18:44] Thank you. Uh thank you so much for your talk.
[18:48] Uh I just wanted to ask you about your failing time uh diagram as a function of um temperature.
[18:54] Namely, what kind of failures did you observe?
[18:57] And uh could you just elaborate a little bit more on that?
[19:00] What kind of failure?
[19:02] Yes, of your Vaxcel's like literal death, or was it just power reduction, or what kind of other failure modes did you observe?
[19:08] Are you mean the the failure of the A cell?
[19:10] Yes, yes, yes.
[19:12] Uh mostly, yeah, there are the several, but uh, you know, long time, that should be kind of power, you know, the reduction.
[19:20] Actually, but uh, you know, we really concerned about the infant failure.
[19:23] So, then, you know, actually, indium gallium arsenide, you know, indium is
[19:27] Gallium arsenide, you know, indium is very good to prevent, you know, dark line effect.
[19:29] So, then, um, yeah.
[19:32] So, then, let me see.
[19:34] They're basically, you know, long time, we have to take care of the, you know, power reduction, anyway.
[19:38] Thank you very much.
[19:40] And let's uh, thank our speaker again.
[19:42] Thank you.