In this episode of the Futurum Tech Webcast – Interview Series, we address the topic of custom ASIC innovation and why it is critical to fulfilling the emerging demands of the global data infrastructure. Marvell’s custom ASIC offering focuses on addressing the intricate, high-speed, high-performance silicon requirements of next generation 5G operators, cloud data centers, enterprises, and automotive applications. Marvell’s growing portfolio of IP at advanced process nodes, now includes 3nm, spanning compute, embedded memories, high-speed SerDes, networking, security, and storage, accelerates time to market and is optimized for performance, power & area, enabling optimal returns on investments.
My guest today is Hugh Durdan, VP ASIC Business Development and Marketing at Marvell Technology, a top-tier semiconductor company. Hugh is a new guest on this show, and shrewdly shares his insights and perspective on the direction of the ASIC market segment and its vital role across data infrastructure environments.
For the foundation of our discussion, we addressed what is ASIC technology, how did Marvell move into the ASIC business, and how the ASIC business fits into Marvell’s overall strategy. From our perspective, Marvell’s ASIC portfolio is key to advancing its industry-leading products across the cloud data center, carrier, enterprise, and automotive markets.
Today, Marvell’s ASIC portfolio uses new 3nm Marvell silicon that is now available for new product designs and includes foundational IP building blocks such as long reach SerDes, PCIe Gen6 PHY, and several standards-based die-to-die interconnect technologies for managing data flow across the data infrastructure. The 3nm development follows multiple 5nm solutions from Marvell – across production or development – that spans its diverse portfolio of custom ASIC programs as well as electro-optics, switch, PHY, compute, 5G baseband, and storage products.
Our conversation drilled down on the following key topics:
- Why 3nm is so important to Marvell’s customers and how Marvell is driving 3nm innovation.
- The industry-wide drive to custom chips and what type of customers and use cases are leading the charge for custom chips.
- The challenges involved in balancing design cost reductions and the need to customize and what is getting customized as customization needs expand beyond speeds and cache size metrics.
- We examined the latest developments in chiplets and Marvell’s view on the ecosystem-wide impact new standards such as UCIe are having on chiplet development.
- What the future of chiplets is looking like and the prospects for an open market for chiplets.
- The prospects for 3D devices and Marvell’s readiness to harness 3D capabilities.
You can watch the video of our conversation here (and subscribe to our YouTube channel while you’re there):
Or you can grab the audio here:
If you’ve not yet subscribed to the Futurum Tech Webcast, hit the ‘subscribe’ button while you’re there and you won’t miss an episode.
Disclaimer: The Futurum Tech Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors and we do not ask that you treat us as such.
Other insights from Futurum Research:
Ron Westfall: Hello everyone and welcome back to another episode of The Futurum Tech webcast. I’m your host, Ron Westfall, research director and senior analyst here at Futurum Research. And I’m really excited about today’s episode because one, Marvell is back and also, I believe we have a very distinguished guest joining us today. And that is Hugh Durdan, the VP of ASIC Business Development and Marketing at Marvell. Hugh, how are you and can you tell us more about your role?
Hugh Durdan: Hey, good morning, Ron. Yeah, its fantastic to be here and thanks for the very nice intro. As you mentioned, my title is VP of Marketing and Business Development for our ASIC business. And basically, what that means is promoting our ASIC brand and capabilities through events like this and also, great deal of interaction with our customers. Marvell has a very focus set of customers that I’ll talk about in a little more detail when we get into the business, and ASIC is really a very much of a co-development with the customer. It’s a joint development and therefore, the relationship with the customer is critically important in the business process.
Ron Westfall: I fully agree and thanks for that background, Hugh. And I think that definitely serves as a springboard to drill down more into ASIC. And I think this is a great opportunity to talk more about what does the ASIC group do at Marvell? And for that matter, how did Marvell move into the ASIC space to begin with?
Hugh Durdan: Ron, good question. So first of all, ASICs, in general, have been around for many, many years. The first real standalone ASIC company was probably LSI Logic that was founded in the early eighties and achieved great success working with, mainly, computer manufacturers back in the day, like the company I worked for, digital equipment and then businesses continue to involve over those many years.
Fundamentally, what ASIC is, like I mentioned, is the co-development. So typically, in the ASIC business, you work with the customer who provides some of their architecture and technology and then we work with them to bring the silicon technology necessary to implement it because typically, the customers are more system companies, they don’t have the experience in the semiconductor design or manufacturing aspects, and yet they have a need for a customized product and that’s why they come to someone like us.
In terms of your question about how Marvell got into this business, Marvell acquired a company called Avera about three years ago, and the legacy of Avera goes back to IBM. If you remember in the late nineties, IBM had always had semiconductor technology internally that they used for their server business. They decided to commercialize that and go into the merchant ASIC business. They were very successful at that for many years and as IBM decided to shift from more of a hardware company to a software company, unless capital focused, they decided to sell off their semiconductor operations to global foundries who took both the manufacturing side of IBM’s semiconductor business as well as the design side. And then a few years after that, global foundries decided that they just wanted to be a pure play foundry. They didn’t want to be in the design business. They didn’t want any perception that they might be competing with their customers. And so they decided to spin out the design portion into a company called Avera that was then acquired by Marvell. And so that’s how we got into the ASIC business.
Ron Westfall: Yeah. I applaud Marvell’s acquisition strategy and I think this is also, with that history background, providing an opportunity to talk more about how does the ASIC business fit into Marvell’s overall strategy, its overall vision. I think folks are definitely interested in that.
Hugh Durdan: Yeah, good question Ron. So yeah, you might look at it and say, okay, Marvell historically has been a standard products company. We do a defined product for a given application, we sell it to many customers. Quite the opposite of the ASIC business, which is a custom product developed with one customer only sold typically to that one customer. So it looks a little bit like apples and orange, but the reality is that it’s incredibly synergistic because what we’ve done is we’ve evolved the ASIC model beyond the traditional ASIC model where the IP and the design was traditionally primarily coming from the customer. And the advantage that we have at Marvell is that since we have all these standard products in different area like network switches and storage and automotive and processors, we’re able to take those capabilities and merge it with the custom model and do a much higher level custom solution than just the traditional ASIC.
And so what Marvell as a company can offer is a full spectrum, going from the traditional standard product where you take something off the shelf and you deploy it in your system if you’re a system company all the way to the other end of traditional ASIC, where it’s primarily the customer’s design and we’re doing the implementation and manufacturing for them. And then a whole range of alternatives in between those two endpoints, where we can take IP from other business units within the company, merge it with the customer’s IP, and put it together in a solution that really lets them differentiate their own product.
Ron Westfall: And I think that shows that there’s a lot going on in the ASIC space and that is certainly something that Marvell’s driving innovation. And in parallel to that, we’re seeing a great deal of conversation about what’s going on with innovation with the three nanometer technology. Can you tell us more about that and why is it so important to Marvell’s customers and partners?
Hugh Durdan: So we recently did an announcement about our three nano meter platform. And the reason why it’s so important is very simply Moore’s law. Even though there’s lots of topic in the press about is Moore law dead or not, I’m not going to get into that because that’s almost a religious argument. But the reality is that there is still a great benefit to our customers from moving to the latest process node. And that advantage comes in the typical dimensions of the power of the device, how much it consumes per function, how much they can fit on a dye, and what the performance that they can achieve from the application. So the types of customers that we’re engaging with are very much on the leading edge in the technology.
It’s things like AI processors, switches, NICs, things that really push the technology boundary either in terms of complexity, memory bandwidth, the performance going on and off the chip, all things that benefit from the latest and greatest process node. So for us developing three nanometer, it’s a corporate wide investment. It’s a technology platform both for the IP and the design methodology that we’ll deploy in our standard products as we develop them in three nanometer and in the ASICs that we do for customers.
Ron Westfall: Yeah. I think that is another example of something that is driving the industry. We’ve talked about what’s going on in ASIC and we’ve talked about three nanometer. Now, let’s talk about custom chips. What is also driving the move to custom chips? What is going on out there that is making it such a high priority?
Hugh Durdan: Yeah, good question. So again, custom chips and ASICs have been around for a long, long time. Decades. But the market has really transformed. In the old days, the primary customers were system companies. So it started with the computer companies like Deck or HP and so on. Then it migrated more towards networking companies like Cisco. And the change that’s happened now is that with the emergence of the mega scales that people like in Amazon or Google or a Facebook, they have such scale in their data centers that they can afford to customize the silicon, not just the system but the actual silicon, to their specific application. So exactly what they need and how they decide to deploy the hardware within their data center is very unique to their particular use case. And sometimes with some of these large companies, you have multiple use cases. If you just take the example of Google, obviously, search is their core business, they also have YouTube, so they have search engines and artificial intelligence applying to that so that when you do the search, the thing that is most likely of interest to you pops up.
They also have video servers as they serve up content on YouTube. And those are very different applications. Each one can benefit from customized silicon. And so the mega scalers have the scale to justify the investment to do the customized silicon and they can benefit greatly because it accelerates whatever their particular use case is, which is going to be slightly different again for each one and not something that they’re going to get from an off the shelf product. So they may start with an off the shelf product, but as they gain traction and as they gain scale, they’ll benefit greatly from a customized product, and that’s where we come in our ability to help customers like that to achieve their success.
Ron Westfall: Yeah, I’m finding that very much is the trend in the industry. Certainly, hyperscalers are playing a major role in driving this across the board innovation. And I think the YouTube example is a very informative one because that allows us to touch again on the customers. What types of customers are really driving the innovation in terms of customization of chips and what types of chips are receiving the most customization priority?
Hugh Durdan: Good question. So first of all, we still do a lot of business with the traditional system OEMs, whether it’s in networking or 5G and so on. The emerging customers are the mega sellers, as I mentioned. The main applications that we serve there in terms of custom silicon as a company, it’s quite a wide range. It’s NICs, so network interface cards, both what’s called a foundational NIC, also smart NICs, which adds some level of acceleration and processor offload in the networking function, we do custom switch chips, we do custom processors, we do custom AI accelerators in and custom video accelerators. Those are probably the main applications that we see today for customization. In terms of emerging markets, one very important for us is automotive. And what we’re seeing there is that, I’m sure you’re experiencing this with your own automobile, if you go back a few generations, very few electronics, maybe antilock braking systems, some engine control and so on.
Now, it’s all about the in cabin experience. And it’s not just the multimedia aspect of do you have a good sound system, it’s the navigation, it’s the collision avoidance, it’s all the bells and whistles that now go into it, which is driving a huge amount of semiconductor content within cars. And in particular, all that driver assistance requires a huge amount of compute power. And so you’re seeing devices go into cars that you never would’ve imagined previously. It’s basically a mini AI processor. It’s not quite as high performance as what goes into the data center, but it’s just still an extremely high performance, relatively speaking, processor that goes into the automobile versus in not too distant past, it was simply eight bit microcontrollers, very simple processors. Now you’re talking about leading edge technology, very complex chips going into the automobile. And in the future, we see that as a large growth opportunity for us.
Ron Westfall: Those are great examples. That reminds me that really Marvell is doing a portfolio-wide drive in terms of making customization something that drives an entire model. It’s not just about ASIC customization, it’s really about the ability to meet customers needs regardless of their situation, regardless, for example, the segment that they’re focused on. So I think this is a sharp differentiator and one I think also allows us to shine the spotlight more on what are the selection criterion. Naturally, through many years speeds and cache size have been important. They continue to be important, but what else is being considered? What else is driving the other offerings out there?
Hugh Durdan: Yeah, good question. So there’s a couple dimensions to it. So let’s start at the very top. The first thing the customer is going to decide is am I going to do a custom product or not? And really, the question you’re asking is why would they make that decision between an off the shelf product and something custom for them? And so there’s two fundamental elements to that decision. One is technical and one is commercial. The commercial one is very simple, Do I have the scale? These designs do cost a lot of money. We’re talking tens and tens of millions of dollars to develop. So they’ve got to get a return on that investment because it is a huge investment. And that’s again where customers like the mega scalers of the world have the scale can get a significant benefit from doing that customization and easily get a return on that investment even though it’s a very large one.
On the technology side, in terms of what they customize and why they customize it. Let me give you an example that I just heard the other day from one of our customers. If you’re a user of Google or you’re a user of Facebook, you don’t want it to go down. You need it to be always available. And so the uptime on the hardware is critically important. And it’s not that a custom device is going to be more reliable than off the shelf device. That’s not the point. The point is that they own the design, they own the hardware, they own the software that runs on it. And so if something goes down and there’s a bug, I’m just making this up, there’s a bug in the firmware that runs on the switch. If it’s their own engineers that wrote that firmware, they have total control about debugging it and getting it back up and running as quickly as possible.
If they had a third party, they bought it off the shelf product, they’d have to go to that third party and say, Hey, I got a problem here. My data center is down. Can you please help? And who knows what’s going on at that other party and where this customer fits within their priorities and how much attention they get. And so they wanted you custom hardware not only to provide the differentiation, the performance and configuration in everything they want, but also to control their own destiny. Another example is a automobile company that very much views the user experience within the car is the big part of their value add. It’s not just what colors did they sell it in and how many square feet of storage space and so on. They really view the user experience as part of their differentiation. In order to own that user experience, they have to own the software and to own the software, they have to own the hardware it runs on. And so they’re developing a custom product just so that they again have total control over their product from top to bottom.
Ron Westfall: And I think these customer conversations are very instructive because it’s pointing to how do you really make the best differentiated product? For example, how does Marvell balance reducing the design costs while also meeting these very unique needs of the customer? How is in essence Marvell delivering the best of both worlds?
Hugh Durdan: And that’s one of the places where Marvell is unique and it comes back to the ability to bring, to bear the intellectual property that we design into our standard products as well as what we have within the ASIC business. So if you look at many of these chips, even though the architecture is unique to the customer and the configuration, so I’ll go back to you mentioned cache, I’ll just… This is a very simple example. A customer may decide that I want a processor complex that has so many cores and such a big cache and these peripherals.
And even though Marvell has a standard product which may fit some of those requirements or the building blocks are the same, doesn’t mean the exact configuration the customer wants. But since we had the building blocks, we can allow them to more efficiently develop that solution custom tailored to their needs than if they just started from scratch. If they just went to ARM or somebody and licensed the raw IP and held to build it all themselves from scratch, it would be a much more involved process. We can actually take the building blocks that we’ve already designed either for other ASIC customers or for our standard products and reassemble them in the configuration the customer wants and greatly increase the efficiency of getting them to market.
Ron Westfall: And yes, speaking of building blocks too, I think that reminds me of what’s also very important that’s developing in the industry and that is a lot of the conversation around chiplets and new standards such as UCIE. And really, I think folks definitely are keen on finding out more about what’s Marvell’s perspective on developments on the chiplet side of things.
Hugh Durdan: So Marvell has utilized chiplets for many, many years. We had a set of switch products which were built out of chiplets, and what that allowed us to do is configure different capacities switches from the same basic building blocks, which was fundamentally a small switch chiplet and a sturdy chiplet that were put together in different configurations in different packages. So what we’re seeing is customers want that same kind of flexibility by assembling different pieces and different configurations to create a portfolio of products because any customer has a range of products they want to maximize the leverage they get out of their development investment. And so if they can build a set of smaller pieces that they can assemble in different configurations and get a portfolio of products rather than getting a single product, that’s much more attractive to them. So that’s one thing that’s driving chiplets.
The other thing that’s driving it is, I’ll call it, a heterogeneous process technology. So a very simple example of that is that if you take a AI processor chip, always wants to be in the latest and greatest process knowing because they just want to put as much on a maximum die size that you possibly can, achieve the highest performance and lowest power. They also have a requirement for a huge amount of memory bandwidth and they can’t integrate that memory in the actual main dye, both because it won’t fit and because the process technology’s not at all the same between a logic process used from the main dye and a DRAM process. So the industry came up with something called HVBM high bandwidth memory about 10 years ago. And what that is is a DRAM chiplet. It’s assembled in the package with the main processor on an interposer and provides the best of both worlds. You get a very high density memory from the DRAM process, you get a very high density AI processor chip from the logic process, you put them together.
And what we’re seeing is now that concept proliferating. Another example is that analog designs don’t really scale very well in the new process technology. 112 gig SRTs and five nanometer is about the same size as 112 gig SRTs and three nanometer. So the effective cost in terms of silicon space is growing as you go from five to three. So it’s better to keep that SRTs in five nanometer. It’s lower cost. It is proven technology, so you de-risk your design and only putting three nanometer the pieces that are really going to benefit from three nanometer. And that’s another example of something that’s leading to a multi-chip heterogeneous solution rather than one big monolithic die. You mentioned UCIE, so I’ll just, if I may, one more comment on that.
What the industry has struggled with is that there hasn’t really been an industry standard for die to die interconnects. We’ve done many in our history, but they’ve been what I would call a closed system. It was all Marvell chiplets all going to one customer. And there hasn’t been a standard for those different die to talk to each other. And what UCIE brings is the prospect that there can be an industry standard so that not all of the components that go into a solution have to come from the same vendor. We could take not just HBMs from a DRAM supplier, but we could take a chiplet for some other function that we don’t have the expertise in, but the customer needs for their solution. With UCIE as a standard interface, we can combine them together within a package and provide a fully integrated solution to the customer.
Ron Westfall: That’s very intriguing. And yes, the evolution of chiplets will continue and perhaps the standard will help drive this to the point where we can build CPUs or XPUs of any nature by basically taking products from different vendors and integrating them into one offering. How do you see that possibility emerging? Is that something that Marvell sees as something on the horizon, that type of scenario?
Hugh Durdan: We are actually already doing that today within Marvell. And so let me explain a little bit what we do. So I mentioned earlier that a couple of the applications that we were focused on were NICs and NICs fall into two broad categories of foundational NICs, which don’t have a lot of intelligence in them, and smart NICs, which add a processor. Again, customers want to leverage their investment. So we have customers that are talking to us about taking a custom NIC that they did with us and taking it with one of our processors and merging them together within one package to create the smart NIC. So they save their development cost because all they have to develop to get a smart NIC is to develop a new package, not a whole new piece of silicon. And so that is a very real application that we’re seeing. And could it evolve in the future to where again, it could be devices from multiple different companies? Yeah. I think things are going there. It’s not quite there yet. There isn’t a broad market for chiplets today, but the trends are definitely going that direction.
Ron Westfall: Yes, I think it’s pointing toward a more open ecosystem approach and I believe that they can be built in the same way that PCs and smartphones are built. It’s just a little more flexibility in terms of multi-vendor integration. And I think looking at the future is always, I think, a great way to wrap up a conversation as stimulating as this one. So we touched on the evolution of chiplets. How about 3D devices? What do you see? What is Marvell looking at in that regard?
Hugh Durdan: That is definitely the next step. So we talked about three nanometer and the importance of the most advanced process node. While packaging is also an absolutely critical piece of technology for us today in 2.5D, we’re taking different die sitting next to each other in a package. And the challenge that we’re running into is space, really. Again, some of these AI processors, you’ve got two radical size die, you’ve got multiple HBM memories sitting around it. The package just becomes huge and that consumes more and more board space so the only way to address that is to go up and so start implementing in the third dimension. The reason 2.5D is called 2.5D is because there’s an active die, the processor memory or whatever, and is sitting on an interposer, which is made of silicon. It’s the same substance, but it’s not active, it’s just wires and there’s no transistors or anything yet.
What 3D is is taking two active devices and putting them on top of each other. And the benefit that that brings is first of all, just physical space. The second thing that it brings is it can bring some architectural advantages, especially to processors. If you look at an AI processor, it’s array, it’s an array of processor elements. Each of those elements has its own memory associated with it.
A very simple use case for 3D is to take that memory and put it on top of the processor instead of next to it. And what that does is it keeps the memory really close to the processor, but it also lets the processors be closer to each other because the latency of each processor talking to each other and the delay of the interconnect is very critical to the performance. So it has architectural advantages, not just space advantages. And so 3D is definitely something that we’re investing in. Working closely with our manufacturing partners to implement. It still has some challenges. Just one simple one that you can think of is if you take two very power hungry devices which are sitting next to each other and you put them on top of each other, that part sitting on top is going to get really hot.
Ron Westfall: No doubt.
Hugh Durdan: You have to have a more efficient way to get the heat out of the device in order to address that. But it’s definitely something that’s coming and I’m sure the industry will solve the challenges associated with it.
Ron Westfall: Yeah, I believe it will unleash tremendous possibilities, particularly with the heat issue being more resolved. And I think that was just an excellent explanation as to what about 3D devices and the innovation we can expect from there. And I think on that high note, I can say thank you everybody for joining us for today’s webcast. We thank you so much, Hugh, for joining us and providing your knowledge and insights, and I think this is going to provide a lot of valuable takeaways and thought for all the folks out there. Good day, everyone.
Hugh Durdan: Thank you, Ron.
Ron is an experienced research expert and analyst, with over 20 years of experience in the digital and IT transformation markets. He is a recognized authority at tracking the evolution of and identifying the key disruptive trends within the service enablement ecosystem, including software and services, infrastructure, 5G/IoT, AI/analytics, security, cloud computing, revenue management, and regulatory issues. Read Full Bio.