Nvidia announced that it would provide full support for its High Performance Computing (HPC) stack to ARM processors, making them full-featured counterparts to the already supported Power and x86 chips. This should be of considerable benefit to companies wishing to embark on high-performance computing with an ARM-based architecture, given the paramount importance that Nvidia has become a high-end computing space.
Companies and their fans like to discuss companies that have achieved the greatest achievements in computer science. However, when it comes to the HPC accelerator market, there is literally Nvidia and everyone.
Seriously. Apart from a 1.5% fragment created for the Xeon Phi 5120D and the "Other" category of 8.2% (yes, mainly other Nvidia cards, although there is a handful of Xeon Phi refugees in this section), the the whole accelerator market HPC is made up of various types of Nvidia GPUs. In recent years, Xeon Phi and even AMD have gained some additional wins in these systems, but both companies have been almost completely drawn from the HPC accelerator for the top 500 systems.
Until now, the entry of ARM in the server space was a weak and shaky thing. Several years ago, Rory Read, CEO of AMD, had predicted that ARM would have 15% of server space by 2018. It's hard to find performance information current ARM server market, but according to IDC. full The non-x86 server market generated revenue of $ 2.1 billion in the fourth quarter of 2018. This represents a decrease of 21% over the same period of the previous year. The x86 server market grew 18.7% in the fourth quarter of 2018, for a business turnover of $ 21.1 billion. With Qualcomm effectively out of the market, the only leading ARM ad in this space was Amazon AWS.
But Nvidia may be thinking long-term with such announcements. There is at least one ARM-based supercomputer in preparation – the post-K system under construction at RIKEN in Japan, with its A64FX architecture, 512-bit SIMD, 7nm FinFET design and HBM2 use .
HPE built last year for supercomputer around ARM processors (although without massive vector units). Cray, which is bought by HPE, has also built one. These machines may not be equipped with GPU accelerators, but ARM is clearly beginning to break into the high-performance computing market. It makes sense for the company to do it.
But Nvidia does not just bring part of its software stack to ARM. The announcement of the company states: "NVIDIA brings to the Arm® ecosystem its full set of artificial intelligence and high-performance computing software (accelerating more than 600 HPC applications and all artificial intelligence environments) The stack includes all NVIDIA CUDA-X AI ™ and HPC libraries, GPU-accelerated artificial intelligence infrastructures and software development tools such as PGI compilers with support. OpenACC and profilers. "
It is not yet known how it will fit into future licensed products, but in theory, ARM server chips could grant a license to the IP needed to connect to Nvidia GPUs. via their own NVLink implementations. This is not to say that Nvidia has made the overwhelming bet on ARM to succeed in its existing investments in x86 and Power. But this represents a substantial "higher level" for ARM, and this could lead to new developments and more products competing with x86 chips in the server space further into the future.