October 28, 2020

Nvidia in Advanced Talks to Buy ARM, Upend Silicon Industry

Those rumors about Nvidia being in talks with SoftBank about purchasing ARM have been upgraded...

Those rumors about Nvidia being in talks with SoftBank about purchasing ARM have been upgraded to “advanced talks.” (Does that make these “advanced rumors?”)

Even if SoftBank can come to an agreement with Nvidia over selling ARM, which it bought for $32B, the regulatory scrutiny from various nations would be enormous, as Bloomberg reports. Apple, Qualcomm, AMD, and Intel all have architecture licenses from ARM, allowing them to design their own CPUs that are compatible with ARM’s instruction sets but that otherwise contain custom IP. Dozens more companies depend on ARM’s extensive hard-IP licenses for various CPU solutions. Given ARM’s ubiquitous position in smartphones, and its burgeoning presence in HPC and servers, everyone from Ampere to MediaTek is going to be concerned about ARM being owned by any single silicon company.

What’s the Advantage of Ownership?

In my previous story, I stated that buying ARM would give Nvidia an easy path to return to desktop and laptop computing with an integrated ARM/Nvidia SoC. What I should’ve addressed then — and didn’t — is how this would be different from Nvidia taking out an architectural license (which it already has), in the first place. After all, Nvidia already builds chips like Project Denver and its successor, Carmel, on an ARM architecture. Owning ARM doesn’t change that.

What owning ARM would do is give Nvidia control over how the entire ARM IP stack evolves in the future. If it wanted to pour development into ARM’s Neoverse server concept and develop new SIMD extensions that would speed its own HPC workloads, it could do so. Instead of being limited to an Nvidia-specific implementation, ARM could design said extensions directly into the standard.

Running multiple Docker container-based demos on Nvidia Jetson Xavier NX

Running multiple Docker container-based demos on Nvidia Jetson Xavier NX.

There are other potential advantages for Nvidia as well. The company could design a low-level GPU as a replacement for ARM’s own efforts, then extend the IP across its core families as well, giving the GeForce brand significant reach across the mobile ecosystem.

Regulatory issues, however, could still scuttle the deal. Historically, Nvidia has always preferred a very closed development model. The company doesn’t license CUDA to anyone and it typically prefers to develop its own value-added software and hardware capabilities as opposed to creating cross-vendor ecosystems. So long as Nvidia is just one ARM licensee among many, this presents no problem. If Nvidia were to buy ARM itself, however, the numerous firms that rely on ARM licenses would demand guarantees that their access to future products or licenses wouldn’t be impeded by anti-competitive measures. If the deal gets to this point, Nvidia will undoubtedly make a number of concessions and guarantees to avoid the appearance of favoritism.

What Nvidia would be buying, with ARM, isn’t just the ability to take out an architectural license. It has one already. What it would be buying, ultimately, is the ability to influence how ARM SoCs evolve in the future at multiple price points and markets. If Nvidia thought it would be useful to their own position to implement CUDA for mobile GPUs, they’d be able to do so. If they wanted to introduce a high-end hard-IP GPU core under the GeForce brand and position the SoC as a gaming solution, they could do that as well.

Just How Shelved Is AMD K12?

One thing I’d love to know is just how far AMD got with K12 before they shelved it and whether the chip might ever see the light of day. According to AMD contacts I spoke to when the company decided to pivot towards Ryzen, the K12 design wasn’t scrapped — AMD just decided that the ecosystem wasn’t mature enough to justify bringing the product to market. The scuttlebutt around K12 always suggested it was similar to Ryzen, with a number of shared design elements between the cores. While ARM and x86 are two different CPU architectures, it would be much easier to cross-leverage IP between ARM and x86 then between, say, x86 and Itanium. There’s no evidence that AMD finished the design or continued to evolve it in the background, but they wouldn’t have thrown the chip away, either. If ARM starts chewing into x86’s market share, I expect AMD might dust off K12, update it for the modern era, and bring it to market.

AMD’s K12 slide. This is most of what we know about the one-time product. AMD has never said how much of the work it completed before shelving the CPU.

Right now, the CPU market is more dynamic than it’s been in decades. A new ARM owner could send major ripples through the company’s long-term trajectory. Intel is struggling with manufacturing issues. AMD is gaining market share. Heck, even open-source efforts like RISC-V continue to drive engagement and interest. Any Nvidia effort to buy ARM can likely be read as an intention to push into x86’s turf in one market or another.

Feature image is Nvidia’s Orin, a self-driving car module with onboard ARM cores and an Ampere-based GPU.

Now Read:

Source link