# CASPA2022 SIGII Advantest SonnyBawari

https://www.youtube.com/watch?v=jDXg3OH_8l8

[00:00] Uh, Bawari, he's the VP of Adventist Cloud Solutions.
[00:05] Uh, which is a part, part of Advantest Corporation.
[00:11] Um, he is a VP of Global Business Development and Operations for Advantaged Cloud Solutions.
[00:19] And he has 30 years of international leadership experience.
[00:25] Going to high positions such as VP, co, and president in a lot of semiconductor manufacturing and product R&D validation area.
[00:36] And he helped develop top results.
[00:40] Delivered, he delivered top results at Intel Corporation where he worked for 20 years and Ang Semi for seven years.
[00:50] And at SGmbH and Adventist.
[00:55] Mr. Sonny Bawari has a Master's in Electrical Engineering from Arizona.
[01:01] state university and uh mba
[01:04] uh mr uh sonia barari's uh benwari's
[01:08] speech title will be hyper edge
[01:11] computing and ai
[01:13] machine learning drive drives high
[01:15] yields and quality of that of advanced
[01:18] packaging for chiplets
[01:20] so without further ado please join me
[01:23] welcoming mr sunny bhavari
[01:33] thank you
[01:34] can you please help me turn my video on
[01:36] it's blocked right now
[01:41] all your needs for plugs okay yeah my
[01:44] video is blocked if you can
[01:49] okay
[01:50] [Music]
[01:54] so you're sharing right now okay yeah
[01:56] i'm sharing i just my video is uh yeah
[01:59] i'll try to get uh your fix your video
[02:02] Okay, no problem.
[02:03] All right, uh, so if you can see my screen.
[02:06] Uh, so I'll start with a brief introduction about Adventists.
[02:10] Uh, we are a global leader in the 80 industry.
[02:13] Uh, overall worldwide, we have 30,000 systems installed.
[02:18] Uh, we have nanotechnology products which support technology down to the one nanometer node.
[02:24] Uh, we have a very diverse workforce.
[02:27] We have 50 countries.
[02:29] We are a top 100 global tech leader, and VLSI Research has awarded us the best supplier for 32 consecutive years.
[02:40] And again, we are actually, if you look at our presence in the semiconductor test, we have over 60 market share.
[02:49] And that's really, uh, something we're proud of and continue to, uh, work on new solutions to enhance our leadership and our contributions to the industry.
[03:00] So this slide, uh...
[03:02] I think so a few uh months ago we had Michael Chan who's uh also at Advantage Cloud Solutions.
[03:08] He presented a lot of the story on what we are doing in this arena.
[03:13] And my goal is really for those of you who probably saw that presentation earlier to take it to the next level to talk about the additional new arenas we are uh deploying and activity we are doing with our customers.
[03:26] So in this slide I think if you look at the background uh you will see uh I think some of you may recognize this is the Tower of Babel.
[03:36] This is a tower uh in the Middle East uh in the Israel Jordan area.
[03:42] Basically this tower is uh humans around 5000 years ago attempted to build a skyscraper which goes to the heavens.
[03:51] So they collaborated on this exercise and their idea was to build this giant tower which takes you all the way to heaven.
[04:00] Now when God heard about it he was obviously not very happy.
[04:04] so what he did was he basically ensured that he dispersed all the people working on this tower to different parts of the world and speak different languages so they could no longer collaborate on this tower to heaven and that's why this tower lies unfinished even after 5000 years.
[04:23] the reason i mentioned this analogy is it's really the same as what we are seeing in the semiconductor industry.
[04:29] if you look at semiconductor manufacturing it's spread throughout the world.
[04:33] our data is in many different continents.
[04:37] people different speak different languages.
[04:39] and not just that i think even the data we have is in very different formats.
[04:44] the data which we are trying to make sense if someone in the back end in the wafer test or in the system test tries to make sense of the e-test or the fab data it's completely different formats completely different types of data continents apart and to put it all together and make sense out of it.
[05:04] is a very daunting task.
[05:06] which baffles the best of our engineers.
[05:09] so that's that's really the situation we see today.
[05:11] and although we are not trying to build that tower to heaven.
[05:13] what we are trying to do is build a tower to semiconductor uh performance habit.
[05:19] which is by bringing all this data together and integrating all these aspects.
[05:23] whether it is chip design or fab processing or test insertions.
[05:29] really bring it all together so we can make sense out of it and analyze it.
[05:33] just like what facebook or amazon does.
[05:37] when you're on facebook you get suggestions you might want to include me as a friend based on past interactions.
[05:42] or you may get on amazon a suggestion that you might want to buy something else while you're buying one thing.
[05:50] so those kind of imagine those kind of possibilities when you apply it at the dye level at the individual semiconductor dye level all the way.
[05:59] and if you know the history of each dye all the way from when it first starts in the fab to the time when two and a half
[06:05] three months later it comes out as a packaged unit or even the system inside your iphone or inside your computer.
[06:13] you can really make a lot of great decisions which improve the performance of your overall supply chain.
[06:19] so so that's essentially what uh what our vision is and what we're trying to do by horizontally integrating all the data and although this one says production test we are actually integrating uh also the the back end test.
[06:33] i think my camera is working now thank you.
[06:37] so uh on the test ecosystem um if you look at what we what the test today is all about is testing is used when you talk about post silicon validation.
[06:48] when you do wat it's also known as e-test when you do wafer sort when you do final test burn-in system level test and all of these we have solutions and advantage has a unique insight into all of these areas.
[07:03] but again what we also see many challenges which our partner
[07:07] our partners and customers have every test insertion has its own set of challenges.
[07:13] uh there's a more and more automation which requires the data to be there at the right time.
[07:18] uh the challenge is in terms of complexity we have known good dye we have system on chip so many dye which are really only available in the dye form coming together as chiplets on a very complex package and you have to somehow guarantee that when the whole package works together on assist on a die on a single package then you have the performance you need even though you didn't get a chance to test each individual individual component.
[07:45] uh also extreme thermal challenges make it a little bit unpredictable.
[07:49] uh the dye which we have tested and tested and wafer sword may actually perform differently so it's really i think point solutions are not the way to go that's the messages if you try to do this with point solutions you you miss the big picture and you may find a bit find yourself with surprises at every step of the way.
[08:13] So next slide.
[08:17] So this talks about a solution which we introduced in our last meeting.
[08:21] When Michael Chang spoke about what our initial focus was on edge computing.
[08:26] Now edge computing is something which, uh, it's actually a lot of you must must know that semiconductor industry provides the AI chips which are used for machine learning by companies like Google and Facebook and so on.
[08:39] But in our own manufacturing and design processes, we are not using it as much.
[08:44] And that's what we started doing with edge computing.
[08:48] The essence of edge computing is we created this inference server.
[08:51] Which basically, uh, try to use my pointer, this inference server over here.
[08:57] It's an X86 architecture and it has a GPU in it.
[08:59] It's connected by an Ethernet, Ethernet link, 10 GB directly to the host controller on the work on the tester.
[09:07] This allows us for near real-time complete performance.
[09:10] So all the computer intensive requirements which happened today on the.
[09:14] tester as you have more and more complexity you have 5g rf.
[09:17] you have nemo.
[09:19] you have multi-site.
[09:21] all of these challenges which are actually the tester is not able to keep up with.
[09:26] because the horse post workstations updated every three four years.
[09:29] you can take advantage of by offloading all the machine competition to the edge server.
[09:34] so that really creates many possibilities.
[09:36] one is your computation and all the algorithms which you have certainly run much faster when you're able to use the data real time.
[09:44] secondly you're able to run applications in their native machine learning languages without transferring them to test touch programming languages.
[09:53] as you know many of our data scientists don't really know test programming languages.
[09:57] they know c plus plus pi torch but they don't know how to write test programs.
[10:01] which is okay because they can do their job independently and create machine learning models independently of what the test program is doing.
[10:10] so this actually allows us to have the best of both worlds have the data scientists keep doing what they do best why the
[10:15] test programmers do what they do best.
[10:19] on top of it this is more secure.
[10:21] if any of you have seen a tester the workstation can be accessed by an operator.
[10:23] and if you're a smart operator you might be able to even hack into it.
[10:28] and whereas this one is a zero trust environment.
[10:31] encrypted security all around and avoid those kind of challenges.
[10:37] but the most important thing i think i want to emphasize is what we have implemented since the last conversation we had with this forum is we have implemented an architecture which is a containerized architecture.
[10:46] which is based on the docker's container.
[10:48] which is an open standard so we were very sensitive.
[10:52] we want our customers to have a completely open transparent ecosystem.
[10:58] so everything we are using the open ecosystem is based on dockers which is an open architecture.
[11:03] what it does is allows our data scientists throughout the world to develop machine learning images which can be downloaded directly to the edge inference server which is near the tester from the cloud and allows you to take the advantage of innovation.
[11:16] anywhere it happens.
[11:17] So today semiconductor companies work with one.
[11:20] There's maybe four or five big analytics companies in the world and many smaller ones.
[11:26] But as you know in our innovation is not the monopoly of one company.
[11:29] There's unique innovations coming from every single company.
[11:33] And what happens is, uh, since a semiconductor company is only working at one time with one analytics company, they are not able to take advantages of the other ecosystem partners.
[11:42] What we have done is we have enabled this kind of app store like product, which we are announcing on July 1st.
[11:47] So the idea is this edge server, if you can imagine, is like an iPhone.
[11:53] Then the container hub is like an app store like product, which allows you to have innovations from multiple ecosystem partners.
[12:01] So you can have five or six options in terms of machine learning innovation, which are available from five or six companies.
[12:08] You can combine with your own in-house data science models and pick and choose what you want.
[12:13] So you are not limited by the.
[12:17] in-house data scientists you have you're not limited to one partner you can have any of these partners and the partner's job is also easier because they don't have to figure out how to work with the tester.
[12:28] all they have to do is develop the machine learning models which go on this inference server and we take care of everything else.
[12:34] so that's really the ecosystem we have introduced and again we are very happy to say that we are seeing tremendous feedback uh from our partners.
[12:43] whether it's partners or customers we have uh tremendous feedback and this is what you will see when we launch this on july 1st is a is a product which has a chance uh just like what you see whether it's app store or google play store you have a chance to select which machine learning models you want to deploy.
[13:05] whether you want to improve your test time so it depends on your pain point do you want to improve your test capacity do you want to get more revenues do you want to improve the quality of your product do you want to improve your yield.
[13:17] depending on what you use cases you can
[13:18] pick the right machine learning.
[13:20] from whichever ecosystem partner is out there.
[13:23] and we have partnership with analytics companies.
[13:26] with ada vendors like synopsis.
[13:29] we are working with fan companies so you can take the data from the fab in terms of metrology which is non-electrical data.
[13:37] and use it to predict what the performance is then test.
[13:40] so the beauty of all this is when you're testing a part at test.
[13:46] you can make real-time decisions on whether that part is good or bad or needs more testing.
[13:52] this is something you could not do before.
[13:55] when i was at intel 30 years ago as a product engineer.
[13:58] i would have to go manually look through a pile of rejects after testing to figure out why they failed.
[14:05] what should what should i do to the test program to either make them pass or figure out whether they are really bad.
[14:11] and i should scrap the whole lot.
[14:13] so those are the kind of decisions people have been doing post processing after the fact.
[14:18] and the possibilities now is that these
[14:20] can be done real time.
[14:21] so you don't have if you go to any semiconductor factory around the world.
[14:25] i guarantee you you will see lots sitting on the floor.
[14:28] waiting for a decision because the engineer does not know they are waiting for someone from the fab to tell them something about the lot.
[14:34] they are waiting for to collect more data.
[14:38] they are waiting to change the test program and these delays piled up on the floor.
[14:40] and imagine this comes at a time when the industry is struggling with supply side shocks.
[14:47] you're not able to ship enough parts.
[14:50] at the same time you have lots sitting on the floor which cannot ship because there's a positive data to make decisions.
[14:56] so this is what we are addressing and the feedback has been unanimously good from our customers.
[15:01] which are all major names in the industry as well as the partners.
[15:05] they all like this concept and uh actually we are in deployment mode.
[15:08] so we have shipped this product the inference server.
[15:11] we have shipped it in the hundreds to top customers just in a year right.
[15:15] so the division we are talking about advantage has been around for only a
[15:20] year and a half.
[15:21] we have shipped this product in the hundreds.
[15:22] our customers are excited and asking for more.
[15:24] and we are continuously adding the machine learning innovations to be available to use to get more value from the edge computing server.
[15:32] so it's really a combination of edge computing, analytics and cloud computing.
[15:36] so one example of an innovation uh if you think about the traditional way what uh uh sunny as an engineer used to do 30 years ago was post processing.
[15:49] which means you would look at the result of a lot a day or so later.
[15:54] in fact i'm being kind when i say minutes, it would take days sometimes to actually make a decision on how to treat the lot.
[16:01] and today you can do real time, which is under 10 milliseconds or near real time.
[16:06] you can make a decision while you're testing a part.
[16:08] should you test it more?
[16:10] should you fail it faster?
[16:11] should you pass it?
[16:12] uh should you test all this?
[16:14] should you consider all the locks on all the die in the lot to be bad?
[16:17] those are decisions now which are enabled near real time.
[16:21] and by integrating our this is one example of a partner pdf solutions where they integrated their solution called extension with acs edge.
[16:29] so by doing that we can get the best of both worlds.
[16:32] what it does is it dynamically adjusts bidding in real time based on the decision engine.
[16:38] and again without any changes to the test program that's the big thing.
[16:40] in the past you would have to change the test program to put this machine learning inside it.
[16:46] and you can use it to improve your downstream operations what does that mean.
[16:50] so downstream if you are doing wafer sort and you want to improve your burning capacity or your test capacity or your test quality you can make those decisions by predicting what it should be based on the wafer sword performance.
[17:05] and that's that's i think the beauty of it and uh in one example with one customer we were able to reduce the burning capacity by 50.
[17:14] and this sounds like a cost initiative but it was not because of burden capacity constraints this customer was not able to ship 50.
[17:21] was actually able to ship 50 more parts because they knew which parts to burn in.
[17:27] which parts to burn in less and so on so those decisions became a lot more intelligent as a consequence of this application.
[17:36] another example i think this is actually a unique approach by a partner we have in israel or proteam techs.
[17:39] what they did was uh they have actually implemented technology where they put on-chip sensors throughout the diet.
[17:48] so these on-chip sensors are collecting data throughout the life of the diet from the time it starts out in the wafer fab to the time it's actually even sitting as a part inside a car inside a tesla or inside any other vehicle.
[18:02] it's continuously collecting data.
[18:04] so what did what the what it does is this in this case as everybody knows uh idtq or crescent current is a good predictor of what the quality or what kind of outliers you can get in a in a lot.
[18:19] so what it does is you basically are able to compare what you expect on iddq.
[18:23] versus the measured results.
[18:25] the machine learning models tell you.
[18:27] okay the measure results are different.
[18:29] than what i expected so it's an outlier.
[18:31] and these highlighted devices make the difference between shipping a part which could eventually fail in an electric car down the road or a regular car versus containing it in your factory and not shipping it and saving a lot of money because everybody knows i think from my experience when you ship a lot defective part to a customer especially in automotive uh and it's uh the claims running easily into the millions of dollars.
[18:55] so it's huge savings for especially in automotive but also in many other ways.
[19:01] so let me talk a little bit about the machine learning model and how does when i talk about machine learning i think many people have many opinions.
[19:08] and yes violet is a great thing right and a lot of companies are saying they are doing machine learning.
[19:13] let's look at the data.
[19:15] according to recent gartner report only for 15 to 20 of data science projects are completed.
[19:21] from those projects which do get completed the ceos say only eight.
[19:25] percent of them get value.
[19:27] so let's multiply the two numbers.
[19:30] the success rate is then only two percent.
[19:33] so imagine a lot of companies are investing in data science projects.
[19:36] there's a lot of press releases but the success rate is only two percent.
[19:41] uh if i was a ceo i would not be happy with the situation.
[19:44] and it would not justify the investment i made.
[19:47] so what do we do in acs we support our customers in getting more value in this journey because it is a journey.
[19:52] i don't want to under oversell it this is something which does take months.
[19:56] building a machine learning model is not a piece of cake takes months but once you do it it saves a lot of money and value.
[20:04] and what we do is if you look at the machine learning ops life cycle.
[20:10] uh there is many aspects to it some of them sound like common sense.
[20:12] so to develop a machine learning model you have to understand the problem which is problem exploration.
[20:19] you have to have the data science and machine learning train the model.
[20:23] uh depending on the use case then you have to deploy it on the floor.
[20:26] so step one two three i think most
[20:28] people
[20:29] do it uh
[20:30] some people better do it better than
[20:32] others but some people are all
[20:34] everybody's trying to do steps one and
[20:36] three
[20:37] where i think people miss the boat is on
[20:39] the monitoring and validation
[20:41] once the model is deployed
[20:43] then people move on to other things
[20:46] and that's really not effective and
[20:47] that's where people uh sometimes miss
[20:50] mr big pictures uh you have to
[20:53] constantly monitor the model the
[20:54] effectiveness of the model
[20:56] detect any changes in the model and if
[20:58] you look at the semiconductor supply
[20:59] chain
[21:00] i think there's so many variables you
[21:02] know everybody here knows i don't have
[21:04] to preach the choir but there's a lot of
[21:06] variables throughout the supply chain
[21:07] which actually lead to this picture
[21:09] which which which has to be comprehended
[21:12] and addressed as part of the
[21:14] part of the machine learning model
[21:16] buildup and this is where i think based
[21:18] on our experience we are adding value
[21:20] and to our to our ecosystem partners and
[21:22] to our customers and by doing this we
[21:24] are able to make sure that the value
[21:27] continues to be there and we don't end
[21:28] up in a situation where the prediction
[21:31] accuracy of the model diminishes over
[21:33] time
[21:35] uh in addition to that i think we
[21:37] announced i'm happy to say that we
[21:39] announced a few weeks ago
[21:41] our program with universities so to us
[21:44] our ecosystem does not stop just with
[21:46] the industry partners
[21:48] we are also opening it up to the
[21:49] university partners we have started with
[21:51] europe with the university of stuttgart
[21:53] and also announced it at recent european
[21:56] test symposium in barcelona
[21:58] we are also working with the us
[22:00] universities as well as
[22:02] uh universities in japan
[22:04] to open this up and
[22:06] our partners get access to big data
[22:09] because one of the key challenges is
[22:10] getting the data so we are providing our
[22:13] partners data which they can use
[22:15] to develop these ai and machine learning
[22:17] models our goal is to not just focus on
[22:19] the next few years but to help have the
[22:21] ecosystem partners the university
[22:24] students of today who graduate tomorrow
[22:26] and are the data centers of the
[22:28] scientists of the future
[22:29] to have the career paths and the
[22:31] motivation to succeed and to understand
[22:34] that they are really playing a big role
[22:36] in creating innovative algorithms across
[22:38] the semiconductor of life change
[22:41] so again this is an example this is one
[22:43] just another slide on what acs
[22:45] university program is it's about giving
[22:48] access to big data to the students
[22:51] uh innovating the new models and
[22:53] applications uh having these solutions
[22:55] then benefit the ecosystem
[22:57] publish these advancements such as you
[22:59] know
[23:01] institutions like cash performances
[23:02] which have the focus on uh on the
[23:05] partnering as well and i think those are
[23:06] things we can continue to partner with
[23:09] to make sure this this really gets the
[23:10] visibility and this is something which
[23:12] benefits the industry overall not just a
[23:15] grand test or any particular company
[23:18] so this is uh i think
[23:21] what i want to talk about in the context
[23:23] of edge computing and cloud computing i
[23:25] also want to spend a few minutes talking
[23:27] about some interesting solutions we have
[23:28] developed which really
[23:31] illustrate
[23:32] the value you get from
[23:34] from advanced test cloud solutions
[23:36] uh one example
[23:38] is uh
[23:39] adaptive pro cleaning
[23:41] now you think pro cards are a very basic
[23:43] thing everybody should know how to clean
[23:44] from cars
[23:46] you'll be surprised that we have talked
[23:47] to many customers big and small
[23:50] and many of those almost without
[23:52] exception
[23:53] we have yet to meet a single customer
[23:55] when we talk about this
[23:56] who tells us
[23:58] no they have it they figure it out so
[24:00] you can see clearly there's still
[24:01] pockets of innovation which have not
[24:03] been addressed even by the biggest
[24:05] companies in the world
[24:06] so what does adaptive proof card
[24:08] cleaning do
[24:10] a simple example is for years as you
[24:12] know the industry has used adaptive
[24:14] tests
[24:15] how do you enable and disable tests but
[24:17] now we have figured out how to do
[24:18] adaptive proclaiming
[24:20] uh now if you look at the historical way
[24:23] uh way the pro cards are clean today
[24:26] there's typically a standard frequency
[24:28] you clean the probe cards at let's say
[24:30] after every once a lot
[24:32] or months after every six wafers and so
[24:34] on so it's a very
[24:36] routine uh almost brainless process
[24:39] and it's actually a big source of
[24:40] quality issues because if you don't
[24:42] clean the pro cargo you have a dirt
[24:46] contamination
[24:47] the contact resistance goes up
[24:49] and you end up with a lot of scrapped
[24:51] wafers
[24:52] which actually as you can imagine right
[24:54] it's not just a yield issue it's a
[24:55] quality also issue also where we end up
[24:57] shipping less everybody ends up shipping
[24:59] less wafers which means less units for
[25:01] transformation
[25:02] so what we have done is we have created
[25:04] a machine learning model where
[25:06] we actually while we are testing the lot
[25:08] we start developing a machine learning
[25:10] model
[25:11] on the very first wafer based on the
[25:14] model we know what the adaptive uh
[25:17] frequency should be so we are using the
[25:19] frequency of
[25:20] what it is what is the actual
[25:22] uh actual actual uh performance
[25:26] suggesting to us in terms of how often
[25:27] the wafer should be cleaned
[25:29] and by doing this we have been able to
[25:30] come up with a
[25:32] very intelligent solution which has
[25:33] produced some amazing results for our
[25:35] customers
[25:36] uh if you look at different types of
[25:38] probe needles whether it's cantilever or
[25:39] vertical
[25:41] we have seen reduction in terms in
[25:42] cleaning which is a maintenance time as
[25:44] well
[25:45] from anywhere from 65 to 85
[25:48] and in some cases we have even seen the
[25:49] yield improve
[25:51] so again this is just some amazing
[25:52] results and we again this is just an
[25:54] example of what you can do with machine
[25:56] learning which wasn't done before
[26:00] we have a how to do predictive
[26:01] maintenance
[26:02] uh as you know today there's many
[26:04] components that test there's eight
[26:06] different tests compost cell components
[26:07] the socket the hardware the tester the
[26:09] handler
[26:10] to bring it all together and to do the
[26:12] maintenance you end up doing it at
[26:14] different frequencies which means the
[26:15] whole system goes down for one part
[26:17] being maintained
[26:18] if you know when the system needs
[26:20] maintain maintenance you can actually
[26:22] synergize
[26:23] and predict the next breakdown and align
[26:25] and get the most value for your
[26:27] maintenance this is the holy grail
[26:29] this is what we are now able to do
[26:31] and lastly i just want to talk about
[26:33] parametric testing
[26:35] a lot of you know e-test which is the
[26:37] last step in the wafer of the wafer in
[26:39] the fab
[26:40] now since this is just a test of the
[26:42] process health of the wafer fab
[26:44] typically out of the thousands of die on
[26:46] a wafer only nine sites are tested
[26:49] when these data for the nine sites is
[26:51] sent to the back end
[26:53] the assembly and test guide the back and
[26:54] says oh it's only nine sites
[26:57] no wonder i don't know what's going on
[26:58] on my on my unit so what we have done is
[27:01] put some intelligence into how we do we
[27:02] test
[27:03] when the nine sites are tested if one of
[27:05] the sites is showing some kind of
[27:07] anomaly
[27:08] we end up testing many more data near it
[27:10] so we get any type of gross failing area
[27:12] gfa we are able to detect it quickly we
[27:15] are able to collect more data
[27:17] so able to fix the problem at the source
[27:19] before it shifts to to the back end
[27:22] so those are some examples i think of
[27:24] what we have accomplished and
[27:26] i think i'm almost at the end of my time
[27:28] and so
[27:30] if it's okay i'll stop here
[27:35] four questions yes
[27:37] okay
[27:39] okay thank you uh um sunny