Nvidia has announced that it will bring full support for its High Performance Computing (HPC) stack to ARM CPUs, making them full peers relative to Power and x86 chips, both of which are already supported. This should be of tremendous benefit to any companies that want to break into HPC with an ARM-based architecture, given how important Nvidia has become to the high-end computing space.
Companies and their fans like to argue over which firms have pulled off the biggest achievements in computing. When it comes to the HPC accelerator market, however, there’s literally Nvidia and everybody else.
Seriously. Apart from a 1.5 percent sliver carved out for the Xeon Phi 5120D and the 8.2 percent “Others” category (itself mostly made up of other Nvidia cards, though there are a handful of Xeon Phi refugees in this section as well), the entire accelerator market in HPC is made up of various flavors of Nvidia GPU. In years past, Xeon Phi and even AMD had a few more wins in these systems, but both firms have been all but driven from the HPC accelerator space as far as the top 500 systems are concerned.
Thus far, ARM’s entry into the server space has been a faint and faltering thing. Years ago, AMD’s then-CEO, Rory Read, predicted ARM would have 15 percent of the server space by 2018. Numbers on the actual performance of the ARM server market are difficult to find, but IDC reports that the entire market for non-x86 servers generated $ 2.1B in revenue in Q4 2018. That’s a 21 percent decline from the same period a year previously. The x86 server market grew 18.7 percent in Q4 2018, at $ 21.1B in revenue. With Qualcomm effectively out of the market, the only high-profile ARM announcement in this space was Amazon AWS.
But Nvidia may be thinking long term with announcements like this. There’s at least one ARM-based supercomputer in the works — the post-K system being built at RIKEN in Japan, with its A64FX architecture, 512-bit wide SIMD, 7nm FinFET design, and use of HBM2.
HPE built a supercomputer around ARM CPUs (albeit without massive vector units) last year. Cray, which is being acquired by HPE, has built one as well. These machines may not be outfitted with GPU accelerators, but ARM is clearly beginning to make inroads in the HPC market. It makes sense that the company would do so.
But Nvidia isn’t just bringing over part of its software stack to ARM. The company’s announcement states: “NVIDIA is making available to the Arm® ecosystem its full stack of AI and HPC software — which accelerates more than 600 HPC applications and all AI frameworks — by year’s end. The stack includes all NVIDIA CUDA-X AI™ and HPC libraries, GPU-accelerated AI frameworks and software development tools such as PGI compilers with OpenACC support and profilers.”
It isn’t clear yet how this will all fit together in upcoming licensed products, but theoretically, ARM server chips could license the necessary IP to connect to Nvidia GPUs via NVLink implementations of their own. This doesn’t mean Nvidia has made an overwhelming bet on ARM to succeed over its existing investments in x86 and Power. But it does represent a substantial ‘level up’ for ARM, and it could lead to further developments and more products competing against x86 chips in the server space farther down the line.