← Back to Library

An interview with mohamed awad about chiplets

While much of the semiconductor industry fixates on raw processing speed, this interview from Chipstrat reveals a more urgent, physical bottleneck: power. The piece argues that the era of simply adding more transistors is over, replaced by a desperate race to maximize performance per watt within rigid energy envelopes. It offers a rare, ground-level look at how the industry is pivoting from monolithic silicon to a modular "chiplet" future, driven not just by innovation, but by the sheer impossibility of powering the next generation of artificial intelligence.

The Power Ceiling

The conversation, featuring Mohamed Awad, General Manager of Arm's Infrastructure Business, immediately reframes the AI boom. Chipstrat reports that the discussion has shifted from "FLOPs or about chips" to "gigawatts of deployed power." This is a critical distinction. The article highlights that the demand for compute is "impossible to quench," yet the physical ability to supply energy is hitting a wall. Awad notes that the industry is now exploring extreme solutions, with people "talking about nuclear power plants and putting data centers in space and underwater," all in an attempt to manage the power constraint.

An interview with mohamed awad about chiplets

The core of the argument is that optimization must now happen across the entire stack, from the software workload down to the silicon itself. The piece illustrates this by contrasting the early days of data centers—"a bunch of off the shelf components in a Stanford dorm room"—with today's reality where giants like Microsoft and Google are building custom racks with "custom networking and custom CPUs." This shift is effective because it grounds the abstract concept of "AI scaling" in the tangible reality of electricity bills and thermal limits. However, critics might note that while the focus on power is accurate, the piece glosses over the immense geopolitical and supply chain fragility required to build these custom ecosystems, which could slow the very optimization it describes.

"When you're in a situation where the technical requirements exceed what is possible with off the shelf technology, people turn to optimization across the entire stack."

The Chiplet Pivot

The most significant technical insight in the piece concerns the move away from monolithic designs. Chipstrat explains that wafer costs are skyrocketing as the industry approaches physical limits, making it economically unviable to build entire chips on the most advanced, expensive nodes. The solution proposed is chiplets: breaking a chip into smaller pieces, using cutting-edge processes only for the logic that needs it, and cheaper, older processes for everything else.

Historically, only vertically integrated giants like AMD and Intel could pull this off. The piece argues that the industry is now desperate to create a multi-vendor ecosystem where any company can participate. Awad explains that Arm has moved from simply licensing intellectual property to providing pre-integrated "Compute Subsystems." This allows partners to skip the tedious work of connecting basic components and focus on differentiation. The article details how this approach has evolved into the "Foundation Chiplet System Architecture," a vendor-agnostic standard recently contributed to the Open Compute Project. This move is strategic; by standardizing how these pieces debug, boot, and communicate, the industry hopes to "amortize some of that cost" and reduce the billion-dollar risk of building advanced silicon.

"It's not just about the dollars, but it's about the time to market and it's about the risk. This is a way that we can spread that load across the ecosystem."

The Turnkey Future

The interview concludes by outlining a new hierarchy of abstraction for chip designers. Chipstrat notes that customers can now choose their level of involvement: they can license discrete IP, take a pre-integrated subsystem, or buy a full chiplet off the shelf. This flexibility is presented as the key to accelerating innovation. Awad suggests that the future lies in an "interoperable chiplet marketplace," though he admits we are "still quite a ways away" from a fully realized version of this.

The piece effectively argues that the days of the "jack of all trades" chip designer are ending, replaced by a modular economy where specialization is paramount. By pre-validating connections between memory controllers, interfaces, and processors, Arm and its partners are lowering the barrier to entry for custom AI hardware. A counterargument worth considering is whether this standardization might stifle the very architectural breakthroughs that come from radical, non-standard designs. If everyone follows the same "Foundation" spec, do we risk converging on a single, sub-optimal path for the next decade?

"Focus on the differentiation. Now, the interesting thing about Arm is that you can come to us and you can get the discrete IP if you want... or you can go get a full on SoC from partners like Nvidia or Ampere."

Bottom Line

The strongest part of this coverage is its unflinching focus on power as the primary constraint of the AI era, correctly identifying that energy efficiency is now more valuable than raw speed. Its biggest vulnerability is the assumption that a multi-vendor chiplet ecosystem can be standardized quickly enough to meet the explosive, immediate demand for AI compute. Readers should watch closely to see if the Open Compute Project's new standard can actually deliver the interoperability it promises before the power wall becomes a hard stop for the industry.

Sources

An interview with mohamed awad about chiplets

by Various · Chipstrat · Read full article

We’ve been pulling on the chiplet thread, lately exploring why chiplets matter and what it takes to build a thriving ecosystem:

The Long Arc of Semiconductor Scaling

Chiplets and the Future of System Design

Building the Chiplet Ecosystem

This week at OCP, Arm made an interesting chiplet announcement: it’s contributing an architecture-neutral, vendor-agnostic chiplet spec called the Foundation Chiplet System Architecture to the Open Compute Project.

To unpack the story behind this, I spoke with Mohamed Awad, GM of Arm’s Infrastructure Business. Our 25-minute conversation covers:

How Arm fits into power-constrained AI datacenters

The state of chiplets beyond vertically integrated designs

Why Arm first created its own chiplet specification

What led to the architecture-neutral Foundation specification

What the Foundation specification defines versus what it leaves open

Why Arm contributed the Foundation spec to the Open Compute Project (OCP)

Keep reading for the lightly edited transcript. Or for you 1.5x junkies, watch / listen below:

An Interview with Arm’s Mohamed Awad.

So Mohamed, let’s start with you. I want to get to your current role, but first, can you share a bit about your background? I saw that you’ve been on the Arduino board, you spent time at Broadcom and you even worked on Zigbee at Ember. So a great mix of compute, connectivity, system design. Tell us more.

MA: Thanks Austin. Great to be chatting with you. I started as a software engineer. I had this concentration in data communication. So I always loved it. For a couple years I worked at those old infrastructure companies – Lucent, Nortel – and did some of that.

I went on to work at a little startup that was working on mesh networking because I thought wireless was cool. This was back in the early 2000s and that took me to Europe for a while and then eventually took me to Broadcom where I got to do all sorts of fun stuff, everything from Bluetooth and Wi-Fi through to GPS and I even started a custom ASIC business there which went on to build custom ASICs for the likes of Apple.

In 2018 I joined Arm, I went back to my infrastructure roots and have gotten into the compute side of things in a big way. It was a great time to join. Obviously Arm has been on a quick trajectory since 2018 within infrastructure and excited to have played my little part ...