According to ARM co-founder Hermann Hauser, Nvidia’s recently unveiled Grace deep learning processor is proof that the company will abuse its market position if it is allowed to acquire ARM. The Telegraph reports a conversation with Hauser in which he claims Nvidia is proposing a “proprietary interface between its Arm server chip, Grace, and their graphics processing unit” and that this risks “locking customers into their products.” Hauser further claims this “clearly shows that they will compete unfairly with other Arm-based server companies such as Amazon and Fujitsu.”
The Telegraph story does not explain why Nvidia’s decision to deploy its own proprietary NVLink constitutes intrinsically unfair competition with Amazon or Fujitsu. To be clear: Nvidia does not currently own ARM and it developed NVLink in 2014, more than six years before it made an offer to purchase ARM. I don’t believe Nvidia has made any official statements regarding the future of NVLink and whether it would offer licenses to the interconnect if it is allowed to purchase ARM. But global regulators have the power to require Nvidia not to allow the bus or GPU technologies it licenses through ARM to languish while preferentially reserving breakthroughs for its own products.
Google, Qualcomm, and a number of other companies have raised concerns regarding whether Nvidia would compete fairly with other companies if allowed to purchase ARM. I’m not pretending those concerns are without merit. Allowing any one company that designs its own custom ARM chips to also be responsible for designing and licensing the Cortex processors that currently power the entire ARM ecosystem is potentially a risk. Even if the deal is allowed to go through, it would be wise to put safeguards in place to protect against abuse.
But here’s my problem with the Telegraph’s article as written: It doesn’t explain why Nvidia’s decision to use its own NVLink with Grace is actually a sign of why or how Nvidia would abuse Amazon or Fujitsu. Presumably, Hauser would rather see Nvidia focus on Compute Express Link 2.0, an Intel-founded standard also supported by ARM, AMD, Fujitsu, Dell, HPE, Lenovo, Inspur, SuperMicro, among others. Nvidia, for the record, is one of the others.
CXL 2.0 uses a fundamentally different architecture than NVLink. As ServeTheHome writes, “With the Grace model, GPUs will have to go to the CPU to access memory. In the CXL 2.0 model, GPUs can directly share memory reducing the need for data movement and copies.” So whether CXL 2.0 or NVLink is better, that’s something customers will be able to see and measure for themselves.
Hauser has called on the UK to take action to purchase ARM in an alliance with GraphCore, another UK company. His arguments rely, at least in part, on the perception that there is a need for the UK to retain ownership of part of the silicon design chain for national security purposes. Intel CEO Pat Gelsinger has been saying somewhat similar things regarding the need for the US and Europe to have a silicon production chain less dependent on Asia, though in a very context. Some Asian companies have expressed alarm at the idea of ARM being owned directly by a US company because it could subject them to Entity List restrictions more directly (ARM is currently owned by SoftBank, a Japanese company).
There are a number of dimensions to this issue that have nothing to do with Nvidia at all. Of the ones that do, there are plenty of valid reasons to be concerned about how Nvidia would treat ARM if it acquires the company. Few of them, however, appear to be in play with respect to Grace, an SoC Nvidia intends to bring to market whether it is allowed to buy ARM or not.
Nvidia has previously promised not to change ARM’s business model and to build a leading AI research center in Cambridge, which will remain the central hub for ARM product development in the future. Talks between the company and various regulatory bodies around the world have not advanced to the point where specific concessions or requirements to win approval have been revealed.
- Google, Microsoft, and Qualcomm Don’t Want Nvidia to Buy ARM
- Microsoft Is Building Its Own ARM CPUs for Servers, Surface PCs
- ARM Co-Founder: Sale to Nvidia Would Be a Disaster