According to comments from Nvidia CEO Jensen Huang during a conference call yesterday, we could see Nvidia-branded CPUs in the future, setting the stage for a new level of competition with Intel and AMD.
Nvidia's announcement yesterday of its blockbuster $40 billion ARM acquisition sent shockwaves through the semiconductor market, and many questions remain about how the two companies can merge their portfolios without alienating or competing with current ARM licensees - particularly when it comes to building CPUs. There's been plenty of speculation that Nvidia won't create its own CPUs at all, instead merely using its stewardship of ARM to foster a more healthy ARM server ecosystem that can leverage Nvidia's GPUs tied together with CUDA and its Mellanox-derived networking stack.
However, during yesterday's briefing, Timothy Prickett Morgan from TheNextPlatform asked Jensen Huang, "Will you actually take an implementation of something like Neoverse first and make an Nvidia-branded CPU to drive it into the data center? Will you actually make the reference chip for those who just want it and actually help them run it?"
"Well, the first of all you've made an amazing observation, which is all three options are possible," Huang responded, "[...] So now with our backing and Arm’s serious backing, the world can stand on that foundation and realize that they can build server CPUs. Now, some people would like to license the cores and build a CPU themselves. Some people may decide to license the cores and ask us to build those CPUs or modify ours."
"It is not possible for one company to build every single version of them," Huang continued, "but we will have the entire network of partners around Arm that can take the architectures we come up with and depending on what's best for them, whether licensing the core, having a semi-custom chip made, or having a chip that we made, any of those any of those options are available. Any of those options are available, we're open for business and we would like the ecosystem to be as rich as possible, with as many options as possible."
Nvidia already builds some ARM-based processors for lower-power applications, but having access to ARM's engineering talent will undoubtedly speed the process of designing custom Nvidia data center chips. The company will also have overall control of the ISA, and it's unclear if Nvidia would be compelled to share all of the future ARM architecture innovations with ARM licensees.
During the call, Huang also said he wants to speed up the Neoverse roadmap to bring innovations to ARM licensees faster. Naturally, it would also be in Nvidia's best interest (at least in the short term) to broaden the ecosystem of ARM server chips, and that would require multiple options from a variety of chipmakers.
Nvidia could also drive Nvidia GPU-specific optimizations like CPU/GPU memory coherence into the ARM architecture, which would then incentivize other chip makers to use Nvidia's GPUs in their solutions. That approach could help solidify Nvidia's position as the premier AI solutions provider in the data center.
Nvidia's aspirations might be even larger than "just" building its own CPUs. Zooming out, pending regulatory approval, Nvidia will have CPUs, GPUs, networking, and DPUs all under one roof. These represent the most fundamental components required to build servers, and it's easy to imagine that Nvidia has a broader plan to eventually design its own vertically-integrated servers with networking and software that ties it all together into what Huang commonly refers to as a "data center-level unit of compute."
There's already several ARM Neoverse-based server chips in the wild, like Graviton 2 from AWS, Altra from Ampere, and Marvell's Cavium ThunderX, but it's unclear how those ARM customers will respond to the Nvidia/ARM merger, and the prospect of competing with the very company they license the architecture from. The impact to a number of other startups, like Nuvia, is also an open question.
One thing is for certain; the landscape of the server market will change drastically if regulators approve the merger. Intel is the dominant server chip provider in the market, but recent 'defections' on both the consumer side (Apple) and server side (AWS) of its businesses already signal a potential weakness that few would have imagined a few years ago. A faster cadence of ARM architecture innovations could certainly pose a threat to Xeon's position, not to mention its Habana/Nervana AI products.
Meanwhile, AMD is steadily growing its market share with its EPYC Rome chips, but a broader ARM lineup could present a challenge. AMD's position as another x86 provider has seemingly slowed the adoption of ARM-based server chips, largely due to the plug-and-play nature of the x86 software ecosystem. Nvidia's focus on enabling the software ecosystem (the company employs more software engineers than hardware engineers) and newly-found access to the vast ARM developer ecosystem could whittle some of that advantage away.
Finally, AMD has a mature lineup of data center GPUs, while Intel is in the latter stages of bringing its own GPUs to market. Having both GPUs and CPUs under one roof gives both companies an advantage over Nvidia, which produces only GPUs, but it now appears that advantage could be short-lived.