In The Arena by TechArena - Arm’s Dong Wei on the End of Monolithic Silicon Design

Episode Date: July 28, 2026

In this episode of The Control Plane sponsored by AMI, recorded live at the OCP EMEA Summit in Barcelona, host Allyson Klein is joined by Dong Wei (Fellow and Lead Standards Architect, Arm) and Colin Brix (AMI) for a deep dive into the chiplet economy, the future of firmware, and why open standards are critical to unlocking the next era of AI infrastructure.

I'm Alison Klein, coming to you from OCP Amia Summit in Barcelona. I have two guests with me that are fantastic, and this should be a fantastic conversation. The first is Dong Wei from Arm. Dong Wei, welcome. Thank you. And the second is Colin Ricks from AMA. Welcome, Colin. Thank you.

Great to be here. So why don't we just get started with introductions? This is the first time that you both have been on the show. Dong Wei, do you want to tell us a little bit about your role at Arm and how Arm is showing up at the Open Compute Conference. Yeah, so I work at Arm in the Architecture and Technology Group. I'm the lead standards architect and fellow. I have always been working in the interception between the hardware and system software,

and making sure the components are tied together and the systems are going to be scaled to support whatever the deployment are necessary. and I also served as the president of the UEFI Forum, which is quite recent. I've been actually in this field about firmware standards for a long time, and I also served in OCP as the steering committee member representing the open platform firmware project. The need for firmware standardization is actually very critical because firmware is now entering into the AI generation as the kind of the orchestrator and the glue and the connecting tissue for making sure the systems would work together and the components are going to work together.

Thank you for that. And Colin, you lead marketing at AMI. First time on the podcast, why don't you give us some background on AMI and your company's role in OCP? Yeah, thanks, Alison. Obviously, AMI has been known for a long time, 40 years in the industry for BIOS and firmware. And I think AMI has played this role behind the scenes, as Dongwei was saying, is like, orchestrating the ecosystem together and providing that glue. We've been a little bit silent. We just make things work.

And I think today, with the complexities of today's data center, all of the inflection points that are happening, it's more and more important for AMI to be front and center because the things that are happening downstream really, you know, need to be addressed upstream. AMI is really looking at the holistic stack, not just from the silicon support, but all the way through to solution and data center IT buildouts. So I think we're excited to be here because obviously Open is a really big

sort of topic today of like how do we all collaborate together on solutions, especially in a really fragmented industry, making sure that we're all coordinated and working together. And I think that's the exciting thing. I'm only at AMI for two months, but if I can say one thing is that excitement, it's palatable in the industry right now. And it's exciting to be at a company that is really being that glue for the industry. I was wondering where to start this episode.

And I was thinking back to the OCP Global Summit last fall. And Arm made a really big splash with some engagements around chiplets. And I think everybody since then has been talking about chipplet fever, like, is this the moment, we're going to have an open-chip-lit economy and see heterogeneous players really rise up to deliver a new capability and a new type of design for semiconductors. Dong, the first question is going to go to you. In your personal view, are we seeing the end of the one-size-fits-all silicon era and really seeing inflection into operators designing bespoke silicon using a multitude of vendors to deliver that silicon?

I think it's a combination of both, right? I think the ability to do the monolithic development of chips are still going to be there. But I think it's facing a lot of challenges in the new AI world where different technologies like CPU, I.O. memory, they may not need the same kind of processing nodes technologies, right? So for cost reasons, for performance reasons, there is a desire to do the designs. like the chiplets, right? But also, you know, it introduced complexity by doing so. Component-wise, yes, you can probably do a little bit more modular designs. But as you're integrating to the SOCs, there is a bit of a challenge in terms of how you actually manage to get all the chiplets to work properly together.

And so I think this is going to be a long journey to first with major companies companies like AWS or MVD, right? They would be the silicon vendors collecting the chiplets together to form the SRC. And Arm is doing, I think last month we introduced the AGI CPU.

We're doing similar things as well. Whoever is doing that integration is going to be in this role of making sure everything is working together. But I think in the longer term, there is certainly this desire to get to this chiplet economy where it could be easier for anyone to utilize the chiplets to form whatever SOCs that they want to have for their purposes.

So in that era, I think you need to make sure the standard interfaces are designed so that these SOCs can work with the chiplets together from different vendors. So the standards work would be very important in that space. And Arm has been leading by contributing this FCSA, the foundational chiplet system architecture to OCP, right? And then, of course, in the industry, there are other efforts as well, you know, UCIE, for example, and a bunch of other activities as well. It's still in the early stages in terms of defining the standards and supporting. And I think it's going to be a challenge. I mean, the analogy I have is comparing this to the PCIE era, right? And it took many decades for the PCI devices to be able to interoperate among each other and with the hosts.

I think a similar kind of a saying would have to happen on the triplet side as well. But it's going to be more difficult because you are talking about the integration of the components into the SOC package, utilising a very expensive manufacturing processes. That's different from the PCIE where individual components are developed separately, and you can easily do pluck vests or things like that, right? But how do you do the pluck fast to a chiplet kind of a component? So these are the challenges that we're going to be facing along the way. You know, it may take a while, but I think there's a huge amount of interest in getting us there.

No, I think that one of the things that I know about Arm is that you fix fragmentation quite a bit within systems. And when I think about the chipplet, you kind of described it. There are the industry standard interfaces that need to be solved so that we can connect chipplets. But even with that solved, there's additional complexity in getting chipplets to operate. How do you see the industry coming together to solve that challenge? Yeah, I think certainly it started from defining how the chiplets are connecting to each other. electrically, physically, right? But I think more importantly, as we move forward, we need to actually look into how we define the standard interfaces for system software, for firmware.

I see firmware eventually will become a glue layer to tie these things together. We need to define things like telemetry, discovery, configuration, and control, right? So you need to make sure all these elements of the triplets are going to be able to be combined together as a system rather than just individual components. I love that analogy of the connective tissue of the triplet economy for firmware. It's funny, you know, in my background in silicon, while I was very aware of firmware because I was working for a silicon company, most of that stuff is very transparent beyond from an operator perspective.

When we look at multi-vender solutions, that firmware becomes much more important. So who actually owns the responsibility to course correct when a system fails in this particular environment? Yeah, so in terms of the owner, like I said, in the short term or the immediate future, I think it's the companies who will tie these chiplets together. They will have to tell the chiplet developers what they are expecting, right? people like us, people like Nvidia and AWS, these kind of companies, they will have to tell the chiplet designers what we are expecting them to be. But that is still coming from a kind of a proprietary requirements to the chiplet vendors, right?

But I think over the longer term, we need to kind of settle down on a standard interface, is not necessarily the implementation side, but the standard interfaces where the telemetry, the action, the discovery can be done. the communication can be done in a standard way so that anyone can go through firmware to combine these modules together. Right. Nicole, and hearing Don's perspective about firmware's importance, what is AMI thinking about its role in addressing the opportunity for the triplet economy? Yeah, I think it has changed quite a bit in these last couple of years of we have to be this intermediate voice of, of looking at the solutions that are being enabled downstream

and making sure that we are projecting that voice of that end customer upstream as well, right? And a lot of times it's a little bit of negotiation and it's a little bit of, hey, we know today you may not see the value of bringing these things together through firmware, but downstream, it's going to make sure that your customer and your customers' customers are being enabled in the right way, right?

And things like building insecurity from the beginning point rather than at the end point is really critical to make sure that, hey, we solve these challenges up front and operators can be up and running without worry, right? And there's so much more things now that we have to also worry about just in terms of the IT buildout, right? Power consumption, thermals, things that really need to be, again, designed in and controlled through a firmware stack. So I think AMI plays an interesting role now. It's not just that silicon enablement partner of the past of when a new platform launches, making sure that the hardware works. It's like, no, even all the way down the stack when the IT buildout happens, like, how is that a being controlled. How is it being used? Just going back a few years in OCP land, we had ultra-ethernet

come out about three years ago and we're talking about different ways to interconnect technologies. We have seen different standards being announced at OCP events. Where are we in this evolution? And what is capable today from a chipplet standpoint? Can you deliver an armed CPU chiplet with an AI accelerator from a different vendor? Are we there yet? And when do you think we would be at that point if that's not true? Like I said, I think if you think about the SOC companies utilizing the chiplets to form the SOC, I think we're already there, right? I offered a few examples. Amher being one of them, we are producing this gigantic CPU and we are using the chiplets. So it is there. And I think the difference now is.

that we're looking at establishing these chiplet standards so that in the future, for anyone who want to combine the CPU with accelerators, with other things together to make sure the designed silicon is going to work for the workloads that it intends to work for, I think it will make life a lot easier. And even before the chiplet thing came about, over the last few years, I actually had done this thing called the CSS, right? So we actually provided the Neovirce CSS so that people, it's not necessarily a chiplet at that point, but people can utilize the verified IP and subsystems to build with the accelerators and other things. Very good examples of success is

all these hypers, whether it's AWS or Microsoft or Google, right? All of them are using this kind of technologies to combine with their versions of accelerators for their workloads. So we have been, in terms of arm, we have been providing the ecosystem, the ability to have this flexibility, right? People can design their own solutions. Now, one thing I think about is reliability. And in a monolithic design, you can control if there's a thermal spike. How do you ensure that processor can be resilient to that?

But what happens when a thermal spike in vendor A's triplet takes out the entire package of vendor B? And is there anything in the firmware that actually helps isolate these events and ensure that the system can continue to run? Yeah, that's certainly, I think, if you think about it, right? So if you treat the triplets as individual component, then certainly when something goes wrong with one component, independently, it wouldn't be. able to address the problem in a very consistent way. So what we need to do is to define, it goes back to my desire to establish the standards, right? So you need to have the way for each of the component to communicate to firmware, the telemetry, the error reporting, the power envelope, and discovery, and so on, so forth. And you need to have a framework defined so that when things go

run in one component, the system would actually be able to react to that. And then through the discovery, through the orchestration, through the control, you can actually redistribute the workloads to different areas. And then you can also orchestrate in a way that you can make sure the components that are not acting correctly are not going to be the real engine that takes on your workloads. Right. So you can redistribute. distribute, but that depends on you have a framework that is standardized and the system can react to the problems, right? And then the firmware would be there to utilize that framework and then take actions accordingly, right? So I think that is something that we really need to work on.

Now, I know a lot of vendors, this is a great example of one of the challenges, because in power and thermal, a lot of the vendors want to keep their algorithms like their secret sauce. So how do you convince them to move to that open foundation that is your vision long term without them feeling like they're giving away their secrets? Yeah, this is something that always comes about, right? So people somehow contrast standardization with the innovations, right? And a lot of people talk to me in the past about, oh, if we do standard behavior everywhere, then that means I cannot innovate. That is absolutely not true.

What we need to do is for our standardization effort, we're more focusing on the communication paths. We're more focused on the foundation, right? We're focusing on the language and how to provide the algorithm to support your innovations. That's something that could be hidden beneath the interface. So the standards should be focused on the, interfaces rather than the implementations. And so, you know, I was talking to someone at the coffee shop

early in the morning. These capsules, right, they can be all different. But that's very difficult because then you would have to buy different machines for these capsules, right? But if the capsules can be defined the same, then you can build different machines to support different kinds of coffee varieties. So that could help you to scale the deployment rather than a vendor locking, right? Yeah, that makes a lot of sense. Yeah, so this is very important for us to achieve, actually. Now, Colin, I know that AMI has a history of working with people across the value chain to deliver firmware, but those have been point engagements historically, typically. How does AMI see your role shifting from a point engagement model into one where you're almost an orchestrator of

solution delivery with the entire value chain. Yeah, I think it's an interesting point that Donway just made. It's the standardization thing doesn't stop innovation. In fact, I would argue not having standards creates less innovation downstream because there are a lot of solutions that we're listening to customers. We're hearing downstream these really unique verticals and applications that some of these design ins are going into. And if we don't listen to them downstream, then we're not building the right solutions to support them in the end state. So I think the innovation happens when we coordinate all of that

together. And then you can really have truly new verticals open up, new opportunities for business. And I think AMI plays a really critical role in that we aren't just that, like I said, that enablement, the silicon enablement partner. We actually have to listen. all the way downstream and figure out, okay, how do we start coordinating these things together? And obviously, standards is a big part of that. Like, how do we make sure that we have one control mechanism that things are able to communicate together, that the telemetry is there and understood and communicated? And I think that is one of the things that I see as a development for AMI of,

we can't just look at a platform at the beginning stage anymore. We really have to look at how it's being utilized. And I think once we do that, that innovation happens, right? Yeah. Now, let's just make an assumption that we get all of this right, that firmware as a connective tissue happens within the chiplet economy. What does this do for acceleration of solutions? And how quickly could a cloud provider actually deliver a custom AI chip to market compared to five years ago?

Yeah, so I think one example I have been hearing is, that if this is not even the chiplet, like I said, this is the CSS. So in the universe CSS that we developed earlier, we've seen a partner that could develop an SOC within 13 months. And compared to the other efforts that are prior to that, that's a saving of 80 engineering years. Wow. That's crazy. That's crazy. Yeah. So I think triplet is going to be even more, right, because if it matures, it could deliver the benefits that we couldn't have done without it. And I think getting back to the previous question, I also wanted to say that having the standards give you the ability to actually do more innovations on top, rather than fighting to reinvent the stuff that everybody has already been addressed before, right?

Now, let's take a bigger picture of OCP. There's a huge focus on sustainability. What do we need to do to achieve carbon transparency at the chipplet level? That's a very interesting question. So I think, again, we need to, as we go into the chipplet designs, we need to have more transparencies into, it is going to be a challenge because we don't necessarily know what each of the chipplet are capable of delivering us. So we need to make some standard framework definitions for the chiplet designers to

let the system, the silicon builders, to be aware of all the metrics that we can understand. Otherwise, we would view the chiplet as a kind of a black box. That probably wouldn't be too good. But on the other hand, chiplet allows us to have the ability to reuse, right? And it's modularity is going to give us that. So I think, but by that nature, it is going to give us the sustainability at the beginning as well. But establishing a framework

so that we can actually see through the metrics of all these chiplets, I think it's going to even help more. I wish I had more time to talk to you guys because I have about 30 more questions to ask, but we are at the end of the episode, so I only have one more. We'll start with Colin.

For the technologies that we've talked about today and the topics that we've talked about today, where would you send them more information to find out about what AMI is doing in this space? Yeah, so we're super excited to be able to participate in OCPMIA. We are also working very closely on creating how do we align together and solve these common challenges? I think the first step is to meet and have these conversations. And then also we are looking at creating a partner. alliance, just a loose collective of partners to go out and market and talk about the things that we're doing, making sure that we're bringing customers' voices front and center. We're creating solutions together, but we're also going out and being able to talk about

those solutions to create more innovation downstream. So I think visit amI.com. We have more information about the different OCP events we're attending. And we also have some, if you're interested in registering as a partner as well on our alliance page. That's awesome. Donkwe, how would you send folks for ARM-based information and all the cool technology you talked about today? Oh, of course.com is where you want to be, right? But I think in addition to that, we are focusing a lot on the industry events as well. OCP in the past OCP Global Summit, as you mentioned earlier, was our premier event that we focus on. And I think this year, this OCP,

The EMEA is becoming more important as well. And we have folks from Arm joining this conference as well. I think moving forward, anywhere you can meet with us, these open venues, D.P. And other industry events, I think we're here to listen. We are. We're here to engage with our partners with the ecosystem to enable us to jointly create the standards and other efforts. Thank you guys. I'm Alison Klein and Dong in college.

Thank you so much for being on the show today. Thank you.