I often opine that NVIDIA needs a data center-class CPU to compete with Intel and AMD, both of whom have used tightly-coupled CPU/GPU technology to win the first three U.S. exascale supercomputer deals. Connecting massive GPUs to fast CPUs over a painfully slow PCIe interface will not meet the needs of the future of supercomputing and AI; the GPU must become a first-class cache-coherent citizen on par with a second CPU socket. To that end, today, NVIDIA announced a deal with Softbank to acquire Arm. But NVIDIA already has an Arm license, so why did it feel that it needed to buy the company? Let’s take a closer look.
The answer is pretty simple: NVIDIA wants to broaden its business, and Arm provides the on-ramp to substantial new markets and business models. SoftBank has made massive investments to allow Arm to penetrate the markets for CPUs in the data center, automotive and the Internet of Things. This deal allows NVIDIA to benefit from Softbank’s efforts, and it also brings a channel and business model that will extend access to NVIDIA IP, such as the Deep Learning Accelerator the company open-sourced a few years ago. In the past, Arm has struggled with AI acceleration, but NVIDIA stands to fix that with its low-level IP and research. Also, it’s worth noting that NVIDIA will be able to license Arm’s vast portfolio of software, accelerating time to market for Arm partners who incorporate AI into new Arm-based platforms.
However, this deal (the largest ever in the semiconductor industry) is not without its potential issues. First, international trade tensions present a challenging regulatory environment. That said, NVIDIA successfully navigated similar issues in its acquisition of Mellanox and is already making moves to do the same with Arm. The company announced plans to build up the Arm Cambridge UK campus into a world-class center for AI expertise, a move that should curry favor within the halls of the UK government. Jensen also shared that NVIDIA will build one of Europe’s largest AI supercomputers in Cambridge to attract researchers and scientists.
Second, it will take time and consistency to assuage Arm’s partners’ potential concerns about licensing technology from such a major competitor. Apple, Qualcomm, AMD and hundreds of other firms will need to see that NVIDIA maintains Arm neutrality. In other words, NVIDIA cannot unduly advantage its engineers with early access to future A designs, nor abuse partners with price increases. Furthermore, the addition of NVIDIA IP to the Arm portfolio would be a helpful demonstration of goodwill toward Arm’s business model and partners.
I must admit that I was skeptical when I first heard the rumors about this acquisition, primarily from a concern about regulatory approval and Arm partner acceptance. But the more I consider the potential benefits to NVIDIA and Arm partners, the more optimistic I become. I do believe that RISC-V will get some lift from this, and time will tell of that position will enable RISC-V to challenge Arm for new designs. For the next 3-4 years, however, I think this move will be successful for all involved (except perhaps for AMD and Intel, which would face direct competition from NVIDIA regardless). It will be very interesting to watch.