Semiconductor Manufacturing and GlobalFoundries: How Essential Chips Drive Tech Innovations beyond 2024

Unlock the secrets behind Global Foundries' ascent to semiconductor supremacy and find out how a colossal $1.5 billion boost from the US CHIPS Act is setting the stage for technological innovation. Our guests, John Carulli, Ken Butler, and Shinji Hioki, join us from the frontlines of chip manufacturing to share their expertise on what it takes to be a pure-play foundry in today's competitive market. 

John holds 7 US Patents. He has over 50 publications in the areas of reliability, test, and process development. He is co-recipient of two Best Paper Awards and two Best Paper Nominations working in close collaboration with university partners. John serves on the organizing or program committees of several conferences including the International Test Conference, VLSI Test Symposium, and European Test Symposium. He is a Senior Member of IEEE. His research interests include product reliability, outlier analysis, machine learning, performance modeling, logic diagnosis, and security.

Ken Butler is a Senior Director of Business Development in the Advantest Cloud Solutions (ACS) data analytics platform group at Advantest.  Prior to that, he worked for more than 36 years at Texas Instruments in DFT and test generation, semiconductor reliability, analog product and test engineering, and data analytics.  Ken has a BS from Oklahoma State University and an MS and PhD from the University of Texas at Austin, all in electrical engineering.  He is a Fellow of the IEEE, a Golden Core member of the IEEE Computer Society, and a Senior Member of the ACM.

Shinji Hioki joined Advantest ACS in 2022, focusing on developing the ACS business in Japan. Prior to Advantest, he served as the ASIC Commodity Manager / Technologist at Tektronix for over five years, where he managed foundries and OSATs for ASIC production. Before his time at Tektronix, Shinji had a substantial 31-year tenure at Intel, where he held various roles in the Quality & Reliability organization, spanning development, high-volume manufacturing production, and supply chain management 

In this vibrant episode we track the transformation of silicon wafer dreams into the tangible powerhouse chips that energize our daily devices, dissecting the interplay of design, production, and the crucial customer relationships that drive the industry forward.

Feel the pulse of the semiconductor world as we tackle the elephant in the room: the pressures of Moore's law and the herculean task of safeguarding tech's most sensitive data. This episode is a call to arms for more transparent collaborations between foundries, assembly, and test operations, with an eye on the future where open data flow might just be the magic ingredient for enhanced yield management and efficiency.

As we wrap up our journey, we turn our attention to the fertile minds of tomorrow's tech leaders. We discuss how acts like the CHIPS Act and mainstream conversations about semiconductors are sparking interest among the youth, encouraging them to pursue careers in this electrifying field. Our conversation weaves through the importance of engaging storytelling in tech education, making semiconductor testing and assembly as thrilling as time travel adventures for the next generation. 

Thanks for tuning in to "Advantest Talks Semi"!

If you enjoyed this episode, we'd love to hear from you! Please take a moment to leave a rating on Apple Podcast. Your feedback helps us improve and reach new listeners.

Don't forget to subscribe and share with your friends. We appreciate your support!

Transcript

Keith Schaub: 0:03

In this episode of Advantest Talk Semi, we delve into the ascent of Global Foundries, which, in a brief span of 14 years, has emerged as the world's third largest foundry, headquartered in upstate New York Global Foundries semiconductors are the unseen backbone of virtually every modern electronic device. Global Foundries specializes in essential chips, from powering the connectivity in mobile phones and vehicles, to the intelligence in smart appliances and the heart of AI technologies. So, I'm joined by John Carulli, Global Foundries fellow leading the Test Development Center for Automotive, Ken Butler, Senior Director of Business Development for Advantest ACS, and Shinji Hioki, Strategic Business Development Director for Advantest ACS, and we're exploring how GlobalFoundries and Advantest are navigating these challenges and opportunities, shaping the future of technology one chip at a time.

Keith Schaub: 1:04

So, John, this morning I actually reviewed the CNBC video with your CEO, Tom Caulfield, and talking about the US Department of Commerce announcing the $1.5 billion in planned direct funding for Global Foundries as part of the US CHIPS Act and the Science Act. So first, congratulations on being the first recipient of some of the funds coming from the CHIPS Act. And what I'd like to start with is what does it mean to be a foundry and what is a pure- play foundry?

John Carulli: 1:42

Thanks Keith. Yeah, that was a lot of work for that submission, and we were very happy to receive that from the US government. It was an excellent recognition for all the work that's been done in that space. It took a lot of effort from our local governments all the way up through state and national and we're on our way at this point. From a pure-play foundry perspective, pure-play really gets to the heart of "that's all we do. At our heart we are a wafer technology company. So, for those that are in the pure-play, the pure really focuses on that to differentiate with some of the other players that are in the ecosystem that may also be a design manufacturer. They may also have other aspects that are outside of just focusing on the wafer technology development.

John Carulli: 2:33

When we think about pure-play from the wafer perspective, we will go and we will receive a design from a fabless company usually, does not need to be a fabless company, but from the person who is actually making the chip. They will go off and they will send us that design collateral, usually in something called a GDS binary format, and that goes into making the reticles or the masks that we'll put onto the machinery in the factory and we will go and build up layer by layer, from the transistor all the way up to the metal pinouts, what this final chip will go off and look like and execute out in the field. When we go off and receive that we have that ecosystem, we have process development kits, we have design kits. We give our customer an environment to go off and build these chips. It will then go out after that and go into an OSAT, most likely. It will go off into an offshore assembly test. They will go off and do redistribution layers, bump layers or other technologies to go and get the electrical signals from the chip into a package and eventually into the outside world.

John Carulli: 3:41

After we go through the OSAT piece of it we will go into board assembly. We will go then into an OEM. In each of the steps along the way everybody is doing some level of testing so that we could all make sure that the product, when it ends out in the field in our end customer's hands, is working as we expect it. So, the end-to-end system is quite complex. It goes across companies, it goes across technologies, it goes across geographies, all so that we could deliver these reliable chips in a dependable way each and every day, seamlessly to our end customers. So that gives you a bit of a flavor from the foundry aspect perspective and the pure- play to differentiate from the design elements that are there.

Keith Schaub: 4:26

Great John, thanks. So, Shinji, you spent quite a bit of time also in a foundry, but a different foundry, so I'd like to hear your perspective on your experiences from being at Intel's foundry.

Shinji Hioki: 4:41

Well, I was more like a customer side of foundry and I'm not in the foundry itself but more like we are purchasing wafers from foundry. That was my supply chain management responsibility. So, our designer actually designed the chips using foundry's PDK or design rule and submit GDS2 file, which is a layout information, to foundry and waiting for the foundry to produce wafers. And once we get the wafer, usually foundry provides e-test data, which is a set of test element measurement value on the scribe line X and Y, scribe line and foundry usually tests those e-test structure to disposition bad wafer to be scrapped, and only good one to be shipped to customer. And customer side we receive those wafers past wafer from foundry and we do the wafer sold testing every active chip locations. And whenever we see some bad yield, we need to analyze correlation between wafer sort result back to e-test result and communicate to wafer foundry to understand if there is anything wrong with this wafer, is there any misprocessing or anything. And sometimes it is very straightforward, very good correlation between e-test to a wafer sort, but often cases it's very difficult to understand the correlation, and so we need lots of help from wafer foundry, and wafer foundry, if we tell them exact failure location, we have an ID lot number they look up the historical record of wafer foundry processing, then they may disclose some information of their equipment condition issues or some other operational related issues.

Shinji Hioki: 6:40

So, my position as a supply chain management organization we are closely working with foundry partners to understand, especially whenever something unexpected happens, yield issues or some qualification failure. Whenever we see a reliability problem we talk to foundry to understand the root cause and et cetera. So, relationship to foundry- to- foundry customer is very important in this case. And once we receive wafer the next thing to do after wafer testing is assembly. Then there is another set of big supply chain coordination required which is receiving key assembly piece parts like package substrate to make flip-chip BGA for instance, and also assembly of those packages and those suppliers who makes substrate and OSAT are not located in US. So, we have lots of coordination work, most likely in Asian countries. Then a final test is also done in Asia in likely cases. So, wafer foundry portion is a so-called like a front-end coordination supply chain. But there are lots of back-end coordination required before chip maker can ship, finish the production and ship to the customer.

Keith Schaub: 8:01

Yeah. So, Shinji, you went over a lot of material there. I want to pivot over to Ken. Ken from your days at Texas Instruments, maybe you can unpack a bit of what Shinji was saying in terms of the front-end sort of responsibilities, where the wafer is processed and then going into test and assembly, which is more on the back-end. So, give us some of your perspective.

Ken Butler: 8:25

John and I both worked together at TI for a number of years and TI is what you would refer to as an IDM, an integrated device manufacturer. So, in the early days of TI's history, 100% of everything was all done in-house, and so they would be responsible for the wafer fabs, they'd be responsible for the design elements, they would be responsible for all the assembly test operations and everything was certainly not under one roof. It was a large multinational company and things were going on all over the world, but everything was executed inside of TI. But John and I witnessed the years where TI made the shift that a lot of companies did to more of a hybrid model, where for the more advanced technologies, those started to be done with foundry partners, such as John works for today, and you would see kind of the shift where in the early days, you know, TI would still be responsible for fully specifying the process technology, so they would design the process technology and the foundry partner would be purely manufacturing. But over time that started to change or whatever, and the process development and the process technology was more of a collaborative effort. And then, depending on this, you know today, if you look at them, in some cases it's still a collaborative model or in some cases they're just buying it off the shelf, so they buy the process technology from the foundry. The foundry does all that work and TI basically just does the design and then it's fabricated by a partner.

Ken Butler: 9:52

Now there's still a lot of technologies that are more in the analog space that TI does still the traditional IDM model where they do all the process development, they do the design and they execute everything. So that's all what we've been talking about is this notion of front-end and then the back-end is the assembly test. So once the wafer leaves the factory after it's been tested, then it goes to assembly test operation. You dice apart the wafer, you put them in packages, go through final tests, depending on the part of maybe other stages, SLT, B urn-In, those kinds of things. But that's all the back-end and they'll use OSAT partners in some cases. But you know, today, even still in TI, there's a lot of that that's done in-house at their own facilities, again mostly in Asia, as you mentioned.

Keith Schaub: 10:34

So, Global Foundries, obviously you guys are front- end. I was reading, John, that I think 10,000 to 30,000 wafers a month or so can be processed by you guys, and every wafer goes through maybe a thousand plus steps. So, I wanted to take a minute and if you could walk the audience through, why are there so many steps and what is the type of equipment that's being used and how it's measured throughout its process? If you could walk us through that a bit.

John Carulli: 11:05

Sure, the system here also is quite complex. It's orchestrated by a manufacturing execution system, an MES system that will have the recipes and what reticles that go with what designs need to be executed in what order. So, there's kind of the master brain in the cloud that orchestrates the logistics on how the wafers will make it through the factory from the time it starts to the time that it exits. There are also some pieces that you don't hear talked about as much that are also super critical. All the incoming materials, all the liquids and gases and reagents that come into the factory, all of those are all instrumented and have chemical analysis done on them and there's all sorts of data that is available on the reagents right from day one. All the feedstock is that up to the specifications that we need it to be. Regardless of the technology level that you're at, atoms count. It doesn't matter if you're on the older ones or the newer ones or anything that's in the middle. You know it's all really precision from beginning to end. So, we have all that on the front- end. You know the wafers will go into the machines.

John Carulli: 12:16

There are various types of machines. We have photolithography machines to print the geometries on the wafers. We'll have etching machines that will remove material. We'll have CMP, which is chemical, mechanical polish that will flatten things out, because all these little height differences, planarized across the wafer, makes a difference lithography-wise. If you're ever focusing on your camera and you're trying to get depth of field, we have to, you know, something in the front or in the back is a little blurry. We have to get everything perfectly flat. So, there's "cleans that do a lot of the rinsing and the cleaning and making sure everything gets removed from the wafer between the steps and particles get removed. So there's a lot of specialty equipment and I'm sure I'm leaving a few out. So, I apologize to my colleagues in the factory if I missed your area, but yeah, a lot of details that go on to go add and remove material so that it'll have all of the transistor and the wire geometries that'll build up the chips.

Keith Schaub: 13:14

I was going to say that's really interesting because I was talking to Chris Miller on the last podcast, and we were talking about his book Chip War on the last podcast. And in his research, he said, ah, you know, if you talk about semiconductors, you know what is making a chip. Think, oh, the fabs make a chip. But if you really look like well then, how is the chip made, you go down this rabbit hole which you've just eloquently articulated, that it's very sophisticated and quite complex. So, I'm going to ask, if you had to guess, like, how many different vendors of equipment do you think are inside your foundry?

John Carulli: 13:56

Oh yeah, so I have no absolute knowledge in this space, but I would guess there are hundreds. You know whether you're talking different types of feedstocks, whether you're talking the different equipment vendors, then there's sub-equipment vendors. So, if you have some piece of equipment, a lot of the process steps that we do are under vacuum. So now you have different types of pumps that are part of these systems that all need to work together and talk to each other in some cases. So, yeah, there's layers and layers of vendors that you know all make a relatively tight knit ecosystem where we all have to know a little bit about. You know who's on the left, who's on the right and how do we all integrate together to get the output from any integrated module, any step that you know goes from one layer to the other.

Keith Schaub: 14:44

Yeah, some economists should sit down and figure out what percent of the global workforce is actually working on semiconductors and developing semiconductor chips. That would be interesting unto itself.

John Carulli: 15:00

We get a little bit of that, I think, Keith, and even you get hints of that in some of the things that we're hearing about factories coming up in Arizona and some of the challenges that are there and some of the challenges I'm bringing up. The factories are making sure that this ecosystem of other vendors that need to be around, if you will, the mothership of the is all all there and in place and have the right size and scale and expertise so that that factory can stay running day to day. None of these factories run in isolation. There's a complex web of support that is needed to make sure that everything is driving the right yield at the right capacity. That's required.

Keith Schaub: 15:38

Okay, so I want to take a step back right now and talk about Global Foundries, history and sort of its story, where it came from and how it emerged and how it focuses today on the essential chips and the strategy behind that. So, if you could take us through that, what does Global Foundries do that's different than other foundries, and what are these essential chips?

John Carulli: 16:03

So Global Foundries came to be around 2009. And from 2009 to 2015, it was a set of acquisitions. It was putting together the capacity and scale and capabilities across a chartered semiconductor which was out of Singapore, the AMD manufacturing arm. So, we think of AMD as a fabless company today. Once upon a time they were an IDM, like Ken was talking about, they did everything from beginning to end. So, the manufacturing arm, and specifically our Dresden facility in Germany, and then IBM Microelectronics, which we have some facilities that are in Burlington, Vermont, and in Malta, New York, where I'm based out of. So, all of those were put together over that period of time between 2009 to 2015. And we were also on Moore's Law, following the next node, the next node of doubling the transistor density every couple of years or so.

John Carulli: 16:59

On 2018, Tom Caulfield became our CEO and there was some transformation conversations and strategic pivot that went on. Due to those transformation conversations at the board level and at the senior leadership level, focus of our unique value in the ecosystem really was on our ability to go drive differentiated technologies for all these essential chips that are out there, the chips that were outside of what we're used to talking about often in the press, which are the GPUs, the CPUs, the network chips, right, all of these very high density chips that need high bandwidth, that are multiprocessor in nature, that are also driving a lot of cloud and compute capabilities. So, while that is absolutely critically important to our overall semiconductor ecosystem, there are all these other chips that sit around those main chips and in some cases, there are other chips that have nothing to do with those chips at all. If you start thinking about, you know Apple watches, or you start thinking about smart meters that are out there in IOT that are monitoring you know our electrical grid or that are monitoring our you know gas at the house or water or other things of that nature, there are a lot of different aspects of chips that are out there that don't rely directly, in their fit form and function as a product, on the big chips that we're often talking about in the industry and the challenges of going down Moore's law down into the small nanometer level and so for all of these chips to be successful in the ecosystem, that's kind of where Global Foundries strengths were overall.

John Carulli: 18:42

When we think about some of the differentiations that we have in our portfolio, it crosses non-volatile memory, RF and millimeter wave, which Advantest has a lot of expertise in testing as well. Silicon photonics power for BCD, for bipolar CMOS, DMOS devices, and for now we hear a lot in the industry for quite some time on GAN or gallium nitride, these wide band gap semiconductors, and then also a specialty capability that we have in global on a fully depleted SOI, our FDX platform, to get to ultra- low power, CMOS, for chips that are battery powered, that need to be, need to be on, and very efficient when they're on and be off and very, very off, not drawing any power you know, when they're in the off state, and so there's those technologies as well. So those are the four or five technologies that make up the portfolio that you know don't necessarily need to always be on the Moore's law treadmill, going down to the small nanometers and add a ton of differentiation for our customers and applications and how they drive value in their specific end market domains.

Keith Schaub: 19:52

Perfect. So, Shinji, let me start with you and then I'll go to Ken the four or five technologies that John just mentioned: The non-volatile memory, the RF, the silicon photonics and the power. Could you comment on what in particular about each of these technologies makes them better candidates for these larger node processes than, say, a GPU or a CPU?

Shinji Hioki: 20:24

Yeah, let me think about analog circuitry. So, unlike the big GPU, CPU, which is CMOS logic circuitry dominated, analog circuitry still uses lots of circuits which really require good pairing or matching to the adjacent transistor to transistor. So, the small variation of slight ratio of P-channel to N-channel or width to length, very tiny fraction of those ratio mismatch, cause big impact to the circuit performance, like a current mirror whatever. So, some differentiation difference input has to be measured very accurately. So, the transistor performance has to match very equally in the narrow region to regions and those kind of situations, shrinking the die geometry by following Moore's law actually make it more harder to match having a perfect match to the adjacent transistor to transistor performance. So, keeping a certain node size or dimension is still important. So, there is not so much benefit of keep shrinking the XY dimensions in that case. Performance actually degrades. So that's one example I can think of for analog.

Keith Schaub: 21:45

And Ken. Any thoughts to add to that?

Ken Butler: 21:48

Yeah, Shinji's example is excellent. I would have said variability. You know, just like he said. Variability gets worse as the smaller and smaller geometries that you get into, and you have to control that because yield is king and you need these things to yield.

Ken Butler: 22:03

And the second thing is that you know when you're building a standalone analog circuit like a power management IC, a PMIC or an ADC, you know, or a DAC or something like that, you're not putting in billions of transistors.

Ken Butler: 22:16

You know you're putting in small numbers of transistors, and so you don't need the economy of scale that you do for advanced technologies, for a modern GPU that has, you know, tens of millions, hundreds of millions of transistors just on one tile alone. And we can think about, as we're going into the chiplet world, if you were trying to build an ADC or a DAC that's going to go into a chiplet-based design and you were building on a very advanced technology, the die size would be so incredibly small that just handling and getting into place would be very difficult. So, you know there's some consideration of that as well, as if I'm going to build up a part out of multiple chiplets, then the analog circuit, small as it is, needs to be large enough that I can reasonably handle it and assemble it along with the other components that are going to go into that package.

Keith Schaub: 23:07

Yeah. So, Ken, you basically just made a very strong position of why chiplets are needed. If you have specific functions or technologies that benefit from different node bases, then by definition, you can't build everything on a single node, and as soon as you have to spread them across different nodes, then you can optimize them more efficiently and thus you have chiplets. I think that is what you're saying. So, John, also interesting from the CNBC interview Tom mentioned that around 70% of these essential chips, I mean we think about CPUs and GPUs, those are in the headlines all the time but 70% of the market is other chips, the essential chips that we're talking about here today, and I just thought that was pretty interesting that you guys are right in the thick of 70% of the industry.

John Carulli: 24:09

Yeah, you go look at those chips, Tom's exactly right. All of those chips that go around. Let's go take a look at a data center case again. It takes a fair amount of power to go drive all of that bandwidth in and out of the IOs. You all test it very well all the time. It takes a lot of power to go off and drive all of that in and out. So, I mean the electrons, you know they're only going to go and get us so far with all these series, interfaces and all the tricks that we do in the analog design. You know, even there to try to get it as power efficient as we can.

John Carulli: 24:44

You know, now we start hearing oh well, what about photonics, right?

John Carulli: 24:52

What about GAN in the PMIX is kind of saying they go on the board, that's around it. The constraints that we hear in the data centers is we can't feed enough power, right? They're whole substations, right, that go into all of this, all these cloud server infrastructures that are out there, and so there are a lot of these other chips that are just as important and critical to go get that end mission where we all wish to be in an economical, environmentally sensitive space, in that data center case, and in other cases it's just to make the chip at all, just to go off and get it into a handheld ultrasound or an ultra- simple case. We wake up every day, we brush our teeth. Maybe it's an automatic toothbrush, there's a little microcontroller in there, that we rely on that goes off and drives all that. Some of these things are so ubiquitous, the word gets over-used but they're really ingrained everywhere in our lives and they aren't CPUs, right, there's not a big GPU, that's in our toothbrush.

Keith Schaub: 25:47

At least not yet, and I hope not soon.

Keith Schaub: 25:55

Okay, so what I want to do now is let's pivot over to the recent International Test Conference. So, we have this International Test Conference once per year and, John, you were there and you presented as part of a panel, and I believe the panel was titled the grand challenges in test, and obviously, Advantest being a major test equipment solution provider, we want to talk about the challenges that you're seeing in the foundry side of business and then how that relates or parlays into the backend assembly and test side of it. So, if you could take a few moments and just give us the highlights from that presentation, that would kick us off.

John Carulli: 26:40

Sure, so Anne Gattiker led that as the moderator, Anne Gattiker from IBM, and modeling it on the National Science Foundation or the National Academy of Engineering. You know, what are these big, big ideas, these big challenges that are out there? We wanted to frame it. She wanted to have us frame it in terms of our area of test. But these big challenges that have not been solved yet. They may be multidisciplinary; they may cross a number of years. They have a high impact on society. What are those problems that are out there, and could we think of those and start to frame those up? And I think it was interesting.

John Carulli: 27:18

The outcome of that audience was "oh, I don't know that we have a lot of grand challenges in test. So that was kind of interesting on how that conversation moved along and I think to a degree I think we all aligned on: "Yeah, it was kind of true. We end up with so much pressure under Moore's law, right, we're doubling these transistors every two years. We're going off and adding all this precision and complexity out there, but we've always been forced to, while these things are doubling, and the costs are going down in other regards. But test, we have to test twice as much every two years and we can't put our price up at all. So, we're kind of under double stress and we're always looking out over the horizon. Is the end result of that trying to go knock down all of these barriers that could have been big inhibitors and we've seemed to have been ahead of the game for our journey for the most part seem to be the consensus.

John Carulli: 28:16

The two areas that came out that may not have been underneath that umbrella, I think one was from Professor Mark Tehranipoor out of University of Florida, who was bringing up the concept of securities and the different aspects of security. We see this on the news now all the time as well. Right, all these different breaches and trying to protect for counterfeits and other things that are out there in our overall supply chain or nefarious actors getting in, and we hear these things about people coming in through the key fobs on cars and being able to steal your car, hijack your car, whatever the case may be, and how do you go off and improve the security systems there as well. So, there's many different cases. Every different end market has something that's going on in that security realm and Mark's been right in the middle of that with a bunch of others and there's not a lot of elegant, great solutions that are there. So that was one thread.

John Carulli: 29:16

The other thread I think that resonated from a personal perspective with everybody, but maybe not so much a technical perspective, so it was kind of on the bubble, was the conversation we're having now, which is the data sharing aspect. We've all run into these problems, as Shinji was talking about at the top of the conversation. It's really important that we understand, across the Fabless Foundry handshake, if things aren't right where we want them to be from time to time, why is that and how do we fix it? Well, if either of you only have half the story, it's kind of hard to understand what the root cause is, and then, you know, get back on to the economic trajectory that we want, which is the yield is up, the supply is there, the capacity is there and it meeting the end needs of the market.

John Carulli: 30:02

For all of us who've worked in the technology area, in the university areas, training our students that want to have access to some of this data so they could go learn about how these pieces fit together, people that are in other parts of the chain, like Ken and Shinji were talking about earlier as well, with OSATs and others, board assembly companies, even the OEMs. We all have little pieces of this understanding but how do you really connect it all together? We've all run into that challenge overall. So, I think that resonated with everybody from what I heard in the conversations that went on, but just not in the way of I think everybody was thinking of it as a technical grand challenge, not so much just a business grand challenge for that particular community. But it was a really good conversation. It was really interesting on how it transformed from the beginning of the panel session to some of the end conversations. It had a fair amount of audience interaction, which was exciting to see.

Keith Schaub: 31:14

So yeah, I want to pivot over to you Ken. Obviously, Advantest has ACS. John mentioned the security aspect of it. So, with test and assembly on the backend, we're doing a lot of things in that arena in order to enable and implement the data used throughout the ecosystem but also data sharing throughout the back end. But to John's point, there isn't a lot happening or at least I'm not aware of a lot that's happening between data being shared or utilized from the fab into test. So, I know both you and Shinji have had some experience with looking at that data, maybe in a microcosm. So, could you talk about that data, how it was used and maybe some of the challenges you see with back-end assembly and test, but also how we could potentially connect that to the foundry?

Ken Butler: 32:07

That's a great question, Keith. So, all three of John, Shinji and myself had the luxurious position of working in some form or fashion in IDMs in the early part of our career, and so if you're wrestling with a yield issue on a product maybe you're in a product group and your part is not yielding the way it's supposed to then yeah, it's not always just super easy, but generally you could get access to the test data to be able to go on and look at it. You could go back to the fab and get information from them about the wafers and how they were processed and what equipment was used and any specific things that are going on with the recipe. And now there's always concerns where people are saying, well, I don't want to air my dirty laundry, but in general, if you needed the information, you could get it. But now what we're seeing and in ACS we're certainly seeing this today, you know, because we're such a disaggregated industry you have a foundry partner, you have a, you have an OSAT partner. You're doing the design, or maybe you're doing a design in conjunction with other people that are providing you IP blocks that come from yet another third party. You may be using analytics from a third party, you know. So these things are integrating from all over the place and you need access to all this information. But now it inevitably crosses company boundaries and potentially country boundaries too.

Ken Butler: 33:19

But we see a surprising, maybe not surprising a lot of interest in people really talking about this notion of data feed forward.

Ken Butler: 33:25

Can I get information from the upstream steps and take advantage of that for processing my material, for debugging it when it's not working well, and ACS, central to what we're trying to do, the things that John talked about, we need to figure out how to securely share that information, get the information moved to where it needs to go, but, you know, protect the information so that the owner's IP is not in some way compromised, but allow downstream users to be able to take advantage of it for their operations. How do we do that securely? What are the physical mechanisms for being able to move that data around? Because it needs to move fairly quickly. You know the latencies need to be as low as possible, depending on how fast the wafers move to final test or from final test into the next step or whenever. So fast data movement, secure data movement, but get everybody what they need in order to be able to do their job more effectively in a completely disaggregated manufacturing club.

Keith Schaub: 34:19

And Shinji, some of your experiences from your days at Intel.

Shinji Hioki: 34:23

Yeah. So even in a small, I mean one-die package, single-die package, the challenge is still there. How to connect wafer fab design information to the final test, that's still challenging. But if you think about the future 2.3, 2. 5D, 3D packaging with many, many chiplets we are talking about 50 or 100 chiplets coming to the same package and having a major EIDL disaster at final test and how to debug those situations without having very good close coordination of all key suppliers' data to the final package house, it will be a nightmare. How to effectively isolate a problem chiplet and how to contain the problem to a minimum scrap quantity than how to quickly backfill those problem chiplet, whatever material to keep production running. So those are huge challenge. So having a good connection and good cooperation of data sharing is really essential to make this big challenge successful.

Keith Schaub: 35:37

So, it's interesting all three of you Shinji, Ken, John - so I talked with Chris about this very topic last podcast, specifically about the 2. 5, 3D and all the chiplets coming in and how are we going to scale that and his thinking was that we're going to have to create standards for the chiplets in order for them all to communicate and exchange data, et cetera, et cetera and I know the Foundry, John, maybe you can talk about this a bit, on your side of things, you have much, let's say, more advanced and mature standards that are in place that guide the way for production, and on the backend, test and assembly, it's sort of not as far along. There are standards that are being put forth, but maybe we just talk about that for a second, like what do you guys think that needs to happen or what is happening from your perspective? John, I'll start with you.

John Carulli: 36:39

Sure, I think inside, yeah, our semiconductor factories, we have a lot of standards and a lot of data feed forward and feed backwards because, again, it's just required to meet those atomic level of controls, the defect understanding of understanding where are their defects as soon as they're there, so that you can go off and mitigate, make it go away, keep it down at the lowest level all the time. All the factory you read in the literature and in the newsreels, right, they're looking at AI and machine learning and you know vision capabilities and on and on. So, all of that is just kind of natural part of I think the foundry evolution. In the assembly test, as Shinji and Ken have also experienced, and you cited, there's not as much maturity. I think it's moving in that direction. I think things for dye level traceability improve there. And then also, how do you do that die level traceability though? But maybe not electrically, because those chiplets, you know, with five- micron micro bumps and you know TSVs or other things, we won't be able to test every little chiplet like we're used to at the interposer level or at the multi- chip within a die level. So, we're going to have to get innovation that is there to understand how do we go off and trace all those pieces, and I think that'll happen and I think there's ideas that are out there amongst various companies to go and address some of these aspects.

John Carulli: 38:23

The key piece, then, for me is how do you go get the data moving back and forth to people right? The part that Ken was saying. What you've experienced in ACS is like, we may have data standards like STDF, that are out there, or RITdb was out there for a little while for just pushing the test data itself back and forth, but then how do you enable it? And what pieces do you enable and who can leverage the data and what's their purpose for using the data, so they don't hit you over the head with it. As some of us have experienced, as we've disaggregated, I can talk about it in terms of IDM a little bit earlier and when you're all in the same economic stack, called a company, you all will give and take at the end of the day, because it's all of your revenue. It's all of your margin, you'll figure out how you're going to share that information regardless of someone, maybe de-optimized so that you could optimize the whole solution out there.

John Carulli: 39:17

Well, that doesn't necessarily work so well when each of you are an independent company that we are, that is across this disaggregated chain that we keep on talking about. So now, what's the incentive to get us all to share that data? So, I think standards are part of it. I think enabling some mechanisms like we hear certificate authority, root of trust and then maybe you've got PoPs or maybe you're injecting keys and things, the security aspects that we've heard Professor talk about over the years. So, I think there's mechanics that we can do better at to implement that will help.

John Carulli: 39:53

But, at the end of the day, how do we trust ourselves from a business perspective, that all of this value that we're all providing can all be distributed in a way that makes sense to everybody? I don't know if we've got that understood yet. To wrap up this piece, I think all of us spend a fair amount of time with our lawyers, as we move on in our careers, getting NDAs signed off, these non-disclosure agreements, so that we have the ability underneath some legal framework so we could share this information back and forth. These are lossy processes. These take time. If you have to react quickly and you've got some yield issue or maybe it's not negative, maybe you want to go, and we need to change our mindset. Looking at things in a positive way, I want to open up opportunities that are out there to go through all these layers that we have today is a bit challenging, to say the least.

Keith Schaub: 40:47

I think it's a bigger challenge than well, we know it's a big challenge, but I'm just not sure how we're going to solve it. Like you said, the incentives that need to be put in place, I was wondering if we could just speculate or sort of just talk here amongst ourselves on what do we think needs to happen in order to incentivize or to get that root of trust that you mentioned in order to share data. I'm just thinking from a personal perspective. If I was asked to share my health data, for example, we're in some medical study or whatever and I'm sharing my health data to get some benefit, but I don't want everyone to see or to know what those medical results are. So, I guess they have found ways to do that in the medical field and maybe we need to leverage and apply their best practices to our industry. But any thoughts on that?

John Carulli: 41:43

I've seen a little bit of that, if I remember correctly, in RITdb, which was going to be, I think, STDF 5 at some point. I'm not sure where it is right now overall in the ecosystem. On deployment, I know there's been a ton of effort and a bunch of cool technologies put into it, but RITdb, from a test data standard perspective, had a lot of those features that were in it that would you allow you to off and put those protections in. You may not want people to see inside your design and some people you may wish to see, that are inside your design and so that you could give them various keys to allow them to see it, maybe at sort, but if they're at assembly test in the field, they couldn't see it. Or they could see it in the field for your end customer, but the people that were in the middle of manufacturing it couldn't see it. So, you could go off and there's technologies and capabilities that are in our ecosystems and in our tools. Maybe they need some improvements here and there, but I think there's been some thought leadership put in in various areas over the years to get some of those underpinnings out there. I think that the challenge for me is really just: Are we going to go off and put some effort to put these pieces in end-to-end and then lay it at the feet of the business teams to say, okay, we've got the infrastructure, we figured this out. I don't think in my opinion, it's necessarily cost prohibitive or even a cost adder. I think it'll be a cost improvement even in some of our traditional roles of just standard yield improvement. It'll make some of the machinery work a little better so we could get faster yield and diagnostic answers along the way. But if we could go get the infrastructure that says, yes, we have active tools that know how to go and protect and allow or not allow via root of trust, certificate authority and other techniques that are out there, then now it becomes fully a business problem. If it becomes fully a business problem, I'm somewhat hopeful.

John Carulli: 43:43

We hear these terms shift left all the time, and we hear our EDA counterparts saying, yes, there's a lot of value to go move all of these things that are costly to find in the field and then do a design iteration on it and roll it out. And how do you go get all of that design capability further and further towards the farthest left of the design definition and a lot of that shift left is all about the economics and it's all about, at the end of the day, time to revenue. How fast can I go get that part through all of the qualification capabilities, so that I know I could go ship it in the field? And I'm hopeful that we could go off as an industry and, a lot of the data that we have is really part of this shift left mindset. We need to get the data that is upstream or downstream depending on I don't know, I guess what your perspective is, and kind of get it up at the source. We need to go get it earlier so that we all understand what the impact and consequences are, so we could go off and make better decisions on day one.

John Carulli: 44:44

I think we do the best as an industry when we're solving our end customer's problems and being proactive about doing it and not waiting until we're in task force mode, because we're very segmented across that disaggregated chain and we're waiting for something to break before. There's enough financial incentive to say, okay, how are we going to collaborate to make it better? And we've all been involved in those types of environments, regardless of where we've sat in the ecosystem. Again, I'm coming at this from a beyond GlobalF oundries view, this is just a wherever we're at in the stack, we all have had these experiences, so how do we go help each other out in that space?

Keith Schaub: 45:24

Yeah, exactly so, Shinji and Ken. Maybe from an Advantest ACS perspective, I mean, a lot of the things that we're talking about are some of the main challenges that we've had to address head on. So, if you could comment on, maybe, what are some of the challenges still are, some that have already been addressed but have led to new challenges I think it would be good for people to hear what's happening in the test side of things as well.

Ken Butler: 45:54

I think John covered a lot of this very well. He talked a lot about a lot of things, but he certainly had a lot to say about the business aspects of this problem. And you know, we probably need and I think you said it, Keith, when you were talking about, there has to be benefit to you for in order for you to share your data and I think you can, that's a general thing that you can say about any of these situations. If a foundry is going to share their data with their design partners or the OSAT activity or whatever, then they need to have some benefit for them coming back. And I don't know if that's financial or what it looks like, but we need to kind of think about, you know, how do people essentially monetize or in some other way incentivize themselves in order to be able to share their information. And then there's just the mechanics, which John also talked about. So, there is this idea of this thing called RITdb. That's out there. There's another competing standard called TIMS and STDF, which John also mentioned, standard test data format.

Ken Butler: 46:56

I think, if I remember the history right. It's almost 40 years old at this point. I think it started its life in 1985 and it's served the industry well in its day, you know. But I think that a lot of us are finding now that it's very antiquated. It's a file-based technology. You know there's inherent latencies that are built into that, and you know ACS is, you know, latency is one of the things that we care about the most and we're trying to come up with mechanisms in order to be able to address those issues. We're looking at more streaming technologies, taking some ideas that we see in RITdb.

Ken Butler: 47:27

Another notion that's intriguing to us is this general modern communication idea or paradigm of a publisher, subscriber, Pub/Sub. Because what we find today again, because you're going to take a piece of information and potentially share it multiple different places. You're going to share it to your edge compute capability. You're going to share it to your local database, your customer's database, you're going to feed forward into another test operation. You have potentially one publisher and multiple subscribers that need to be able to get these data streams. You have to be able to do that securely, and John touched on that too, about the idea that when you build these Pub/Sub models, you have to do it where you have essentially authorized subscribers that can get to the data and anybody else who's not authorized cannot.

Ken Butler: 48:11

So, these are all the kinds of ideas that we need to build into a general solution that can serve these modern needs but kind of get us away from the old fashioned way of doing things. But we also find that you know all of our customers and the companies that we all work for, you know they have decades of effort built into the existing data movement capabilities they have today that are still based on STDF for the most part, and you know moving away from that is painful because you have a lot of infrastructure that you have to replace and so that's coming along slowly. But you know people are, you know, now starting to kick the tires on some of these modern technologies to see, okay, well, maybe it's time to think about, you know, doing it a different way, because there are some inherent advantages that we don't have doing things the old way.

Keith Schaub: 48:52

Perfect and Shinji, now you've had time to collect your thoughts.

Shinji Hioki: 48:56

Yes. So, I was thinking about how to incentivize, you know how to give more incentive, and I think it's coming from how to set the goals, how to measure the success for each supply chain. So historically, foundry, it keeps shipping wafers as long as e-test data meets the criteria. So, e-test means only nine points per 12-inch wafer and only those nine points are the criteria to make the wafer shippable or not.

Shinji Hioki: 49:28

But if we think that today's major changes in package assembly from single monolithic die to a multiple chiplet integrated in 2. 5D, 3D in the future, I think interaction with other chips are very invisible, very difficult to predict just by looking at the single die by die. And so, the criteria for success probably needs to be rethought or redefined, looking at the more holistic view, how the particular wafer is consumed by 2. 5D, 3D assembly company and, after they integrate, how my wafer becomes successfully yielding at final product level, if there is something else besides e-test or historical way of delivering data is needed by this final test environment, probably that area we need to figure out how to create a nice incentive to encourage data sharing from FAB and pay additional adequate price for the data if needed. And so those business negotiations or figuring out how to coordinate all this flow without having too much negative security concern and business concern. That's the key, I think.

Keith Schaub: 50:53

So, I think what I heard just to summarize is e-test worked. We would test nine points on a wafer and just from those nine points you would decide if it's good or bad. Even if we decided to dramatically increase e-test, the challenges with the chiplet paradigm are going to force us to rethink how we make measurements and how we use the data throughout the entire chain. So even if we e-tested the entire wafer, it wouldn't be sufficient. So, we need to rethink how we're doing it.

John Carulli: 51:27

Yeah, the transistors are too few in that regard and Ken and I had some experience in this in the past as well with just performance mashing, even at the system level, if you do an independent chip, a singular chip, it's a high-performance analog chip and it's going to go into somebody at the system level, right. There are many times where they go. I don't want to know what the data sheet is that says it's between A and B. I want to know exactly what the population is, because it makes a difference for my precision analog board application that I know exactly where that came from. Well, all of our factories, regardless of whether we're doing sort data or e-test data, say we're between A and B and that's what we guarantee. And it could be anywhere in the middle and there's a distribution and this is what will work, and maybe in days going by, and even a lot of applications. Now that's okay.

John Carulli: 52:21

As Shinji brings up, we go off into the chiplet, especially since you know at least SOC right. We go and we put all this design effort int o see how all those little IP blocks work within the same technology, within the same square frame. Now we're going to disaggregate it, we're going to put it across multiple companies, across multiple technologies. And you know there's times, even today, when you put chip A and chip B on board C, right, that they're all in spec and it doesn't yield. Why? It ends up being a board assembly problem. In that regard, and we hear a bit at ITC from time to time when our board assembly people come on in and give us some of their challenges and, yeah, you just can't parametrically match the way you would like to. And once we find out the certain ones that like to play with each other, then our board yields go back up again.

John Carulli: 53:12

And so this understanding of this richness of the data, the parametric responses Shinji's bringing up especially, goes back to the beginning of the conversation. In an analog world, right, where it's all about the performance and the parametric number, less about a digital it works or it doesn't work type of view of the world, then all of these things start to become really important in a different way of looking at how do we share that data? So, we understand again none of us are happy unless the end product is working. And if the end product isn't working and the customers aren't buying it, then nobody down that value chain is getting anything out of it. So, how do we figure out how we're going to go off and distribute all that in a way that makes sense, because we all want to go make product that right every day, all the time, the first time.

Keith Schaub: 54:02

You know, semiconductor companies in general, the average age is a bit older than other companies and, as you guys have talked about for the last hour, becoming an expert in any one of these challenges can take a decade or more, sometimes two decades, and so that's kind of one of the reasons that we do have, let's say, an older workforce, because it takes so long to get good at it. And then, once you're good at it, you know people want to keep you around. But we have all these new challenges that are coming and that are already here, and we need, you know, talented engineers coming into the field. So, I wanted to give you each a chance to sort of give me your one minute pitch on why should engineers and scientists want to study engineering and come and work in the semiconductor field? So, who would like to start with that?

John Carulli: 55:03

And not go to an AI company and make $300,000 a year out of their freshly minted PhD?

Keith Schaub: 55:08

Well, I don't know. If you told me that I would say I want to go to the AI company. So, yeah what do we do to attract new people, new talent, into this field? Because it's obviously challenging. It's obviously very interesting and exciting. So, yeah, take your minute and give your pitch as to why they should consider that.

Ken Butler: 55:33

I think we wrestled with this a bit when we were still at TI in the sense that, and we talked to some of our strategic universities that we wanted to recruit from them, and there was this constant question about yeah, well, what's the salary look like in the semiconductor field versus what you could get for a hyperscaler or an AI company or something like that?

Ken Butler: 55:54

But I think that you know, like you said, Keith, the problems and the challenges that we're coming up with, that we're coming up against, are so, so interesting.

Ken Butler: 56:04

If we in the semiconductor industry could just do a better job of communicating to the young people out there to say, all these great technologies like AI and hyperscalers and those kinds of things and, like Chris Miller says in your earlier podcast, it's all underlined by semiconductors, everything and there's so many unique and interesting challenges to kind of make these things work in all these different applications, the sky's the limit, and so you just need to broaden your thinking beyond just,

Ken Butler: 56:33

you know, I'm going to be one of a team of hundreds of people developing the next big LLM, and maybe I'm going to work on a smaller team. I'm going to focus on developing a cool new semiconductor product that solves some interesting problem that nobody's thought about. So, I think it's really as much as anything, it's just trying to get out there and help people understand the connection between these two things. And I think that maybe the side benefit of the CHIPS Act and all that's going on in that space is going to help get the word out, because now you know President Biden is talking a lot about semiconductors and you know who knows? Four or five years ago maybe nobody else was, but now they are.

Shinji Hioki: 57:12

Okay.

Shinji Hioki: 57:13

So, to develop the future workforce who comes to semiconductor or technology areas, I think we need to think about the kids, elementary school kids, how they feel or how they appreciate the work semiconductor industry is providing for.

Shinji Hioki: 57:30

And so, lots of the kids understand who are the star players in basketball, baseball, those sports, American football but we do not probably market ourselves very well to the audience for young generations how cool technology is changing the world. So, the GTC NVIDIA, you know they are talking about. It's almost like a rock concert. There are so many audiences but those are adult audience. But we need to have some event to really express how those semiconductor companies, brand new technology is changing the world and who are the heroes of those semiconductor industry, how they grew up, what they studied, and try to create some so-called dreams, dream jobs for the kids, so that kids can think about semiconductor industry can be the choice instead of just going to American football or basketball as a choice for future dreams. So, I really like to have some effort for encourage kids to think for the technology as a future job and education, and those conversations are probably starting point.

Keith Schaub: 58:49

Yeah, Shinji, I like that idea a lot. I think you just came up with the title for your program. You call it "Semi dreams". Sounded like you're gonna start a new non-profit organization called Semi Dreams when you retire. But okay, over to you, John. What do you think?

John Carulli: 59:07

So, there's nothing like the experience of actually getting your hands dirty and working on a problem where it's going to turn into something tangible, a product that is out there. And I think the kids start to see that when, I think you know many of us have gone to the high schools and the elementary schools and we brought in the chips and we brought in wafers and you know I'll have the kids, you know, break the wafer and you could see it cleave. And you know they start being inquisitive about asking things, looking at things, seeing how they actually work out there, and you say, hey, this chip, that's in your phone or that's in your gaming toy or that's in your car, they're all over the place and here's how they look and here's how they feel and, these are really complex and interesting problems to solve and you know, in that phase of their lives they love solving problems. You know, all the way up through college going back to our data sharing, if we would just share some of the data with them so they could get their hands dirty and start playing with these different things and understand that, yeah, you know, they're actually part of this. T hey could stay with the sports analogy, right? This massive team, that all has to orchestrate together to go get these really cool gadgets that are out there that are changing the world in all these different ways.

John Carulli: 1:00:22

A lot of them, right, you know they're very tied to the environment or to other causes that they care about and you know it's kind of hard to look at only any

John Carulli: 1:00:32

causes out there that don't have some sort of technological interaction with it, so they don't have to be a bystander or not get involved. They could go, get in right at the roots and if they're science minded, they could go and start actually making things and making an impact right away. To Ken's point, you know there's always going to be the economics, we're always going to have that. But you know as much as everybody would love to be a billionaire, I don't know that everybody necessarily wants all the different things that go along with that. Are you really living your most fulfilled life overall? Maybe for some, yes, have fun, go do that. But there's a lot of other things that we could do make a good living and do good in our communities, and a lot of that has to do with just getting your hands dirty and go fixing things and making stuff.

Ken Butler: 1:01:20

Keith, I have to interject and tell a short story here, and a lot of that has to do with getting your hands dirty and go fixing things and making stuff. Going back to this idea of trying to get young people excited about this technology. We all of us here on the call are familiar with our colleague, Philip Nye, who worked at IBM for many years, and I can remember him telling me a story about when his kids were grade school age and they had the traditional parents come in and talk about their work. Phil did a lot of things at IBM, but among those things was he was involved with burn-in studies for their ASICs that they were producing, and so when he went and talked to his kids' grade school class, he brought in a handful of chips or whatever and he said I want to tell you what I do for a living.

Ken Butler: 1:01:55

I'm working with time machines, because I take these brand- new devices and I put them in this time machine for a few hours and when they come out they are very, very old and you could see the little kid's eyes light up or whatever when he was kind of telling his story or whatever. We need creative thinking like that about people to go in and come up with clever ways, but relatable ways, to explain this to young people in a way that kind of gets them excited and jazzed about it.

Keith Schaub: 1:02:21

Good story and when you said time machine I got interested. So, I was like: What time machine? Tell me more. Okay, so that brings us to the end of today's episode. I'd like to thank John, Ken and Shinji for taking the time today to bring their experiences of the foundry of test and assembly of front end and back end. Thank you, guys, for attending and educating the community on Advantest Talk Semi.

Ken Butler: 1:02:53

Thanks, Keith. Thank you, Keith. Appreciate the opportunity to come join you.

Shinji Hioki: 1:02:57

Thank you very much.