The Chiplet Revolution Will Be Standardized: The New Building Blocks of Computing

Brian: [00] So Mohit, welcome. Thanks for joining us today. How are you doing? Oh, thank you, Brian. Thanks for having me here. Great to be talking to you again. Before we dive into all things Alphawave and all things chiplets, let’s do a little background check here. So you’ve done, if I have this right, you’ve done ST, Infineon, Rambus, Cy5, Open5, Which was acquired by Alphawave, and now you’re at Alphawave.

Tell us what you do and what your role is.

Mohit: Sure, yes that has been my journey on the semiconductor side. And I’ve been pretty much an IP and a custom silicon guy throughout my life. So currently I manage the custom silicon and IP business at Alphawave. The executive team here and Chiplet is one of the vectors I’m driving at Alphawave.

Yes, I was part of the acquisition Alphawave made about roughly two years ago, which was a carve out from the previous [01:00] company. Lots of my experiences, initial part of my years was developing IP, so I’ve lived that side of the world. Then moved into a lot of the, what I call is the product marketing, application engine roles.

Then eventually grow into the management roles, driving businesses, and building lots and lots of custom chips for our customers. It has been an exciting journey, the journey overall, and I must say, living in a golden era of semiconductors, at some point it felt like a dull moment. Software was taking over everybody after apps, but nowadays everybody is after accelerators, so love it.

Brian:

Brian: Let’s talk about Alphawave, founded about six years ago, went public 2021 came into the world as a SerDes solution provider, if I understand this correctly. Talk about the evolution of the company and from SerDes into much more complex technologies.

Mohit: Definitely. It has been a, I would say quite fulfilling journey.

And yes, the founders of these company, of this company had a You know, they’ve done this many times, built 30s companies, starting at Snowbush, they did vSemi, and then did Alphawave. And I think it, from a get go, the vision was very clear. Built IP, which was the DNA, but used that to grow into the chips and the chiplets world.

And that’s where it brought a lot of attention to me, to when I was in the cloud. other side of the acquisition that this will be a perfect marriage. Because if you look at [03:00] any successful semiconductor company out there, you need some fundamental technologies, right? Which alpha we carry and bringing in the custom silicon, some of the other die to die stuff we were doing.

So it became a very good bridge to build on top of it. Went IPO in 2021 at great valuation, used that capital to acquire companies to fulfill that vision. And proud to say that we are delivering on that vision today. Of course, transforming an IP company into a semiconductor company is not a trivial task.

It might seem easy, that change the name from IP to Alphawave Semi takes a whole lot, right? Whereas once you’re delivering an IP, which is just part of a component of the chip, versus you’re going to deliver a fully package tested solution. Where you need a lot of, silicon capability, operations, supply chain management, all the goodness you need to build actual physical products.

So yes it has been a pretty interesting journey. [04:00] We laid out this strategy about two years ago and pretty much we have been executing to our strategy and we just see that this whole advent of AI has accelerated our strategy even further.

Brian: What was the tipping point for expanding the strategy from IP into a much more complex and interesting world of chiplets?

Mohit: Yeah, so I think as we looked at building lots and lots of IP, we could see one underlying trend building newer complex select and then all the IP we are building is for what I call as enterprise markets, right? Whether it’s networking, communications. accelerators, compute devices, storage.

So we are not so much into the IoT side of the world, but we are building more complex stuff and we could see an underlying current that building large, complex silicon using these IPs is getting tough, whether you can call it because of the process node complexity, the cost of building [05:00] these things are getting crazy.

Nodes are great. TSMC is doing fantastic jobs. I’m saying he’s catching up with Intel’s and everyone is in the race. A job to put a complex silicon out there is still getting tough, right? So that’s where we caught into that there will be an inflection point in the industry that you will have to disaggregate, right?

Over the years when I started my career at Infineon, the whole job was to aggregate stuff. We were building application processors from mobile phones. So there used to be discrete analog parts which were brought in. There used to be these other interfaces. They kept integrating, but I think we have reached a point where we have started to disaggregate just because whether it’s from a cost perspective, time to market perspective, or you can say even risk perspective, it’s not feasible to just keep on going homogeneous or a monolithic solution.

Brian: As you made that evolution, once you made the [06:00] decision to expand into that world, I can imagine that resourcing the engineering department. Was non trivial, right? I have to assume that you needed some new skill sets, some new perspectives.

Mohit: Correct. Correct. No, definitely. Just to give you by numbers, we, when, two and a half years ago, Alphawave was about roughly shy of 200 people.

We have more than 1000 people today, right? Holy mackerel. In two years? Yes, and there’s of course some inorganic parts were there when we brought in the acquisition, but still I feel we hired more than what we acquired. So we have been growing. We recently opened a brand new facility in Bangalore, 500 people office, and we just opened at San Jose office, a bigger facility.

And been expanding worldwide in Toronto and Israel. And Taiwan. So yes, talent is not easily available, [07:00]especially the stuff we are doing. So we are going after places where we can find the right talent, right? It’s just like similar to, what Arm has in pockets of the world, right? The CPU expertise, the Cambridge is the, what I call it, the mecca of the processor, right?

Similarly, Sardis mecca is Toronto because of the University of Toronto, there is good expertise in Bay Area. There are some pockets in other parts of U. S., right? So that’s what we have been after. And then when we brought in the silicon side, that’s where, scaling in India, leveraging Taiwan, and because of the natural reasons being found, Rhianosat’s been there, that added to our capability.

So we have been very Careful very selective where we should go so that we can arm ourselves with the right set of skill sets and the appropriate talent to strengthen ourselves as a company.

Brian: Now, you’ve done some chiplets already. You’ve done the AlphaChip 1600, which is a reconfigurable SerDes.

[08:00] You’ve done the AlphaChip 1200, which is high speed DDR5 memory controller. And now, you have the Arm Compute Chiplet, which is based on Arm Neoverse. Describe the design of the Arm Compute Chiplet and the complexity surrounding that, because that sounds like a much more complex Device if you will.

Mohit: So yeah, we, historically you’ve picked up where you’re at, that you’ve been doing a lot more what I call as connectivity chiplets.

Now we are getting into what I call as a compute chiplets, right? And that’s one of the things we as a team came together that as a part of our initiative to go, grow beyond connectivity and get into compute. This is the third vector we opened up in our chiplet portfolio. So earlier in the June time frame, we announced a partnership with Arm that we are announcing what I call as the industry’s most versatile, [09:00] reconfigurable, and optimized chiplet portfolio, which includes our I O chiplet, where basically you have industry standard UCI on one side and PCIe CXL Ethernet on the other side.

Then you have the memory chiplets, where you can use UCI one side and have DDR or LPDDR on the other side. And now we brought in the Arm compute element, which actually is altogether another orthogonal vector. So the way I think simply, Brian, about this is, like using a human body anatomy, is the compute is like your brain, and connectivity is like your spine.

A controlling our system, which is you used to connect, right? So you need both. So that’s the simplest way I can explain that, we brought in we had a lot of the spinal connections to using the connectivity chiplets, but we needed that computational stuff to bring in. And this allows us quite a [10:00] bit to help customers build for some very, certain interesting use cases.

To deploy with accelerators and other markets, wireless and others we came here after to expand our time as well, right? So that’s naturally does that. And as well as allows us to even build more complex custom silicon with our customers. We are adding, more teams for this. Like I said, it’s a new orthogonal vector.

So that is also one of the areas we have been expanding as a company to bring a lot more I would say SOC level expertise from a computational standpoint.

Brian: So let’s play a little mind game here. Notwithstanding the fact that AI and ML as a workload or as workloads didn’t exist. If you went back to the beginning of your career and you were trying to implement this functionality, it would have been on a very large motherboard, right?

Not [11:00] just, not a chiplet, correct?

Mohit: It will be a, yes, what we call a Lego block approach, right? So it will be a discrete piece of packaged SoC. Sitting there, you will have IO connectivity somewhere else on the motherboard. They’re talking to each other, but, I must say nobody, I’m sure Jensen says the same thing.

It’s not like anybody saw AI coming. But many people happen to be at the right place at the right time. Yep. Ultimately, there are some fundamental things here. The need, how much data as a human we are processing every day, many devices we are carrying. There is inherent need to have these devices, these things be more compact and processing lot more bandwidth at much lower latency and much lower power consumption.

So if you put these three elements together, that makes everyone think, Hey, I need to think about my architecture differently, right? So a lot of that evolution which has [12:00] happened I think is around this. I think clearly, like everyone says, after the Internet era, right? This is the new biggest inflection point we’re gonna see.

Probably in my career, your career, right? After this, God knows what’s gonna come out. But at least we’re gonna live through this 8th generation AI era for the next couple of decades.

Brian: Thanks. So why Arm? You could have built a custom device, no?

Mohit: Yeah, no, I think a very interesting scenario there I’ll play for you.

When we look at the market, you want to solve the problem for where you want to be successful, right? We see Arm as the most dominant architecture out there for the markets we are in. Like I said early on, we are not targeting consumer, IOT, or those kind of market segments. We are very specific into these enterprise class game where when we surveyed, the one of the outcomes [13:00] was, yes, is the dominant force, is it because of preexisting stuff?

Yes. Yes. Yes. But also the software ecosystem, right? The newest platforms has these SVE extensions built in there. So the ease of use for us, it clearly stood out that this is the way to go for us if you want to build compute chiplets. Because we are not trying to be a CPU company, I’m trying to build an SoC compute chiplet, right?

So I want to pick up the processor. Which gets me to that Silicon realization. And so yes, we, and on the other side, you can always say we are going in the direction where the wind is flowing also. Yeah, so that’s also the case. And we did a lot of survey around this, believe me, this was a big thing.

And early on, actually last year I started on this and it was my very personal initiative I took on in my organization that, Hey, We need to drive that partnership with Arm. We have been [14:00] doing some stuff with Arm over either with customer designs, we were working on certain other cortex generations for certain customers, but then when we position ourselves as a company focused into the segments, Neoverse just stood out there.

Hey, you have to be on the side of Neoverse. So it was not a difficult decision from that perspective.

Brian: And you’re targeting 6G applications and or 5G applications as well. Do I understand that correctly? That’s correct.

Mohit: Of course, it’s a versatile, compute chiplet from that perspective.

But yes, one of the use cases prominent use cases is the 6G, which is another in the wireless infrastructure space, a big move, right? There are 5G use cases out there as it’s still being deployed in certain countries, right? And again, all of these things are feeding back from the bigger underlying thing about bandwidth, latency, [15:00] power optimizations from all aspects, right?

Whether a user Accesses cloud on a mobile device, on an edge, using the network, or a enterprise class. They all boil down to the fact that you want much more efficient systems from all aspects. Yes, so one of the very prominent use cases of the compute chip that we are developing is to target the 6G market, which is.

If you look at it, pretty much in the next few years, it will be talk of the town. Yeah, people are already talking about it in the standardization and the organizational bodies. So yes, we want to be ready with that hardware infrastructure alongside on leveraging other software infrastructure.

To be able to play a role in that

Brian: so talk to me a little bit about the architecture of the chiplet yeah, it has onboard radio and pre processing and What else have you integrated?

Mohit: So right now the plan is to we are building [16:00] one compute block which will have interfaces of basing PCIe on DDR5 And we are putting our UCI in there, which allows us to connect the RF elements into it.

So we are not integrating RF directly into this chiplet because, the ADC DACs, you, again, goes back to the disaggregation theory, right? You don’t want them to be on the bleeding edge processor process nodes because those are highly custom analog blocks. So those will be available as chiplets.

And this will be the kind of the compute element of the DPU element to connect to those, right? So this whole Neoverse N3 based chiplet is targeted about the teraops per watt, driving that efficiency for those particular use cases. And the UCIE interface in there allows it to be versatile to connect to our other chiplets or our customer chiplets.

[17:00] It can actually also connect to certain accelerator chiplets, because that’s where a lot of people are going to try to do L1, L2 acceleration with their networks on the phase, so that can connect to another specific piece of silicon. So this is how it can enable what I call as the new terminology is chiplet enabled custom silicon, right?

I like that. So a lot of people say custom silicon, but I think we should start thinking that in future, most of the custom silicon will be developed using chiplets, where certain pieces, I want to be able to just get it off the shelf, if I use that word, and then I have a custom component, which I go design it based on my application, my differentiation to connect it in the package.

As a company we are by the way spending a lot of effort on the packaging side to make it happen also because that’s another area of innovation.

Brian: If you can describe it, what’s the differential in [18:00] complexity between designing something like this for 5G application Versus designing it for the coming 6G era.

Mohit: Yeah, so I think one big difference is like when 5G was old, and I think it’s like you are doing an assisted AI there. A lot of people are trying to slap on AI, the existing 5G hardware, which was built a year or two ago, to make it, smarter and continue till 6G comes out. 6G natively is going to be supporting AI from a get go.

So if you look at any wireless infrastructure, you have mechanics of processing, encoding, decoding, transmitter, receivers in there. I think with this, what you will be able to do with a certain accelerator, you can have almost the neural transmitter receiver. That’s a very big term, and I, I’m just projecting it, but I think that’s where eventually maybe 7G will be like [19:00] that, where these devices on the wireless side will be also able to understand and learn on their own, rather than, tapping into the network back and forth for the workloads.

So that’s one of the biggest thing what people call is assisted AI versus natively AI. So 6G will be a native AI. That’s how the standardization is happening. A lot of work in RAN is around this. So that is going to drive. I think quite a bit of use cases around this and we are trying to work with partners.

So we are not trying to be a full blown solution ourselves. We want to use it to go to these wireless infrastructure partners to put together the entire solution.

Brian: Talk to us a little bit about the value in Neoverse CSS from Arm. How does that make your engineering, your development process more efficient?

Mohit: Oh, yes. That’s one of the, whoever came up with that CSS on Arm [20:00] site kudos to Arm on getting there. I think this is, I think another level of abstraction Arm has created, right, critical pieces of, to create a entire compute subsystem. Arm is able to offer as a system in itself rather than me spending my engineers to sort out those things.

Arm is adding value on top of their IP to build these subsystems, right? So I think it’s a beautiful concept. I think will stay in a lot of relevance. And I saw Arm expanded it beyond Neoverse also to the other families of product lines earlier this year. So which totally makes sense because think of me as an SoC company, right?

I want to build chips. I would prefer my engineers work on the differentiation that I add as an SoC company. So my differentiation to put an even subsystem together is not a problem. And I’m taking that role and building these compute subsystems and giving those out totally makes sense. Any [21:00] company will buy into that fact.

And so are we, that the CSS really helps us accelerate the time to market over here. And I can focus on my resources. To like all these integration aspects, silicon aspects, packaging aspects, rather than figuring out, okay, how should I optimize this break with the code, right? So it takes away a lot of my real dollars, which I would have been spending as on my OpEx, CapEx, right?

To rather move to the other. Differentiating elements for us as a company.

Brian: So talk about your special sauce your value add when your sales team goes out into the field and they’re pitching the technology. What’s the differentiator that seals the deal for them?

Mohit: Yeah, so the biggest thing is the ability to, I would say, customize and versatility of these platforms, right? We are not locking anybody down with, [22:00] hey, this is the only architecture you can do with these chiplets. The, when you, that’s the biggest thing. I know many people say chiplets, yeah, packaging adds cost, it’s complex, I need to understand, redo my architecture.

Yeah. Take your three year roadmap and see what you can do with a chiplet, and suddenly you will realize how much value that brings, and that comes down to the versatility of these chiplets, that you can build multiple skills. I was talking to one customer very recently, and I went in a meeting projecting one of the use cases, but when we made a whole pitch around chiplets, he said, Wow, I can do another product actually within six months if you can give me this chiplet.

So that just opens up people that how much more they can do using the same set of resources, but leveraging the chiplet approach. Yes, it might look like initially you have to put a little bit extra effort to package [23:00]them, qualify them and get there, but it’s still far cheaper than building a new SSE from scratch for every use case.

So that’s the big thing and I would say that’s the biggest sell we do to the customers that don’t look at a point solution or a point solution for your, in your roadmap that I’m solving this year’s Dart in your roadmap. Think of a three year window, what you can do. Similarly, like we used to have these discussions with our IoT, like you talk to switching, they typically build.

50 terabits, which 100 terabits, which people are talking about 200 terabits, which in future, the best thing is disaggregate IOs so that your fabric can keep evolving with better power and area, but you’re not need to be burdened by your service to be available in two nanometer or anything. Is that problem to us to give you as a chiplet?

So same concept, very similar concept. I think a lot of people You know, it’s first there [24:00] was a phase of the chasm we call for any technology adoption. Crossing the chasm, yeah. Yeah, so that’s what’s going on right now. Of course, the big companies like NVIDIA and AMD, they built their own, but it’s a closed box system.

They’re not selling chiplets in the market. The moment the other UCI played the role, hey, I can be a big industry standard language that you can bring in these, I think that chasm starts to cross over. Yeah. And companies like us who know how to package them, where, get them the yields out of it and show the use cases, I think it’s going to bring a lot of new players in the market with us.

Brian: That’s an excellent outlet. So we’ve talked about your value add, the value that you bring to your customers. We’ve talked about. Neoverse CSS, how that’s helped the team out, helped the development effort out. Alphawave also was a founding member of the Arm Total Design Program. Talk about the value you see there [25:00] in terms of the larger context of speeding development efficiency.

Mohit: So as part of the ATDE, and lots of interesting conversations within ATDE. I think like I said, bring in concept to leverage compute subsystems. We as a company were very new, right? Like I said, last year itself, according to this, but that whole total design concept to uplevel the abstraction layer helped my team to get into this journey.

But I don’t think I would have been here if that concept wasn’t in place. So we took in, some of the Arm prebuilt compute subsystems. We were able to verify them, we were able to take them through our design flows, look at the power area efficiency numbers very quickly, because there were reference scripts, so I don’t need to invent the wheel again and again, right?

So that brought in, a lot of confidence also in my engineering team, because a lot of people were nervous. Because, hey, [26:00] How it’s going to work out. We are just new and you’re crazy. GM has thought about building something on this already in a short period of time. So no, but like I said, after once they saw what it, what Arm has been doing and is hoping to get there faster, I think everybody bought into this journey and I think we were quite pleased with all the efforts Arm is putting out there on the total design platform.

Brian: Engineers are by nature conservative and risk averse, but it’s the GM’s job to pull them out of their comfort zone, isn’t it?

Mohit: Exactly. Yeah, and we people play bets, yes, and the fact is somewhere you have to look a little bit ahead, right? And say, okay and you have to bet on the right players.

So that’s why partnerships are very important in the ecosystem. Nice. One other thing about Total Design is, it’s not just in that inclusive or just the SOC layer. [27:00] There are people who are building software as part of that ecosystem. Foundries are there, right? Arm is taking a kind of a very, broader, comprehensive look, what all should be part of this total design platform, because if you have just one of the element and not the other, you may still not have the full solution stack.

So another thing which comes together very well, like I’m introduced us to one of the software partners working with ODNs, how these things are put together in the boxes and how you can do all the other things. So we have been, I would say I’m very pleased by all the effort Arm has made in this regard.

Brian: Let’s pull out the crystal ball. All right. You’ve been around the industry long enough. You can do this in your sleep. Where do you see the future of chiplet technology evolving in the next couple of years? Because And no doubt it’s going to evolve quickly. And what’s next for Alphawave on your roadmap?

Mohit: I’ll tell you there is a funny thing. And [28:00] many people don’t Actually, it was an eye opener. I wrote a article about chiplets in 2017. Maybe I was a little bit ahead of my time. You were. Yeah. The people are like, yeah And I talked about io chiplets at that time, right?

That’s funny. But anyway, I’m very happy that this is coming. That’s why I feel very personally attached to this thing. I’ve been in this for long enough to feel it that and the way it’s headed is in the right direction. Now, if I put my, thinking cap and see what, I think at some point, it’s not there today.

But I think at some point, maybe three years from now, let’s just taking my judgment there, there will be a marketplace where these chiplets, you will be able to go to a marketplace and buy them as non good diets. There will be enough standardization work, which will be done. There will be enough packaging interoperability, which are all the concerns today.

So those will be sorted out. But that’s what standards are [29:00] for. So I think UCI will become one of those essential technologies that just what PCI was in the era we grew up where motherboards were made. PCI is the interface to connect to. And, once you connect over PCI, two chips will talk to each other.

I think UCI will take the same place for two chiplets to be interoperable and talk to each other freely. And you should be able to source those chiplets. I hope, as one of my elder daughters, hopefully she gets into the college. She gets to buy these chiplets and puts complex systems together.

And she doesn’t need to worry, how do I make one chiplet at a time.

Right. So let’s see. I think that’s my kind of vision that this gets to that point. And we want as Alphawave to be one of our key players that we are able to offer in partners like Arm, our own technologies built on top of it.

Be able to be a provider in that place for chiplets [30:00] and people can come to us to package them as well. So it’s not just about getting known good guys, ultimately you need to get system. So that’s something we, we do.

Brian: So you didn’t write that article six years ago for semiconductor engineering by any chance, did you?

I think so. Because Ed Sperling was banging the drum even before that on chiplets and most people thought. He was barking at the moon.

Mohit: Yes, that’s correct. Look where we

Brian: are now. Yes, exactly. Alright, last question for you. Since you guys are a startup that’s now gone public, there are a lot of people that want to get into this space on the hardware side, software side, AI, ML, of course, is catnip to investors.

What advice would you give to companies in the semiconductor space who want to innovate in chiplet technology, get into this space, develop new [31:00] approaches?

Mohit: I think one very key thing for people who are starting and looking to innovate here is think of just, focus on your differentiation. There will be ways other what I call as foundational or generic stuff, which you will be able to get in a form of chiplet.

To get your first proof of concept or first product out there because let’s say you want to get to a fundraise or get your product out. Don’t get into solving everything which has been already put out there. It will not add any value or differentiation. It might add, yes, the proud moment as an engineer that I did that, but honestly, keep the focus on your differentiation and the niche.

Which will actually get you to success faster, right? So because I see a lot of companies come by and say, Hey, I’m going to put this from scratch. And we go in education tour [32:00] with them, with the founders to tell them, Hey, can’t you just focus on this particular area? Maybe we’ll get you leave the rest generic stuff to us that players in the industry and you will get there faster.

A simple analogy I give to many people and since I was in physical IP business back then in artisan days and all that, earlier when you were building chips the foundation cells and memory compilers, these, they used to be called as foundation blocks the chip, I think this is again in my crystal ball brain, I’m thinking that chiplets will take, replace, that you will call chiplets as foundational elements, And the abstraction will shift at a package level where rather than taking these cells and memories to as a foundation cells, you will be actually getting foundational chiplets and you will package them in the, in the system and that’s how you will build the new chip into the hood.

So again, time will tell how wrong [33:00] and right I am, but, let’s see, I’m really optimistic by nature we are optimistic guys but I believe in this technology to art and as a company also, I think we are very committed in this technology across the board.

Brian: We will keep checking in with you and we will keep reviewing the scorecard on your crystal balling as the years progress.

Mohit, thank you so much for your time today. This has been a delightful conversation and very insightful. So thank you.

Mohit: Oh, thank you, Brian. Enjoying the conversation very much on my side as well.