At ISC19, NVIDIA is announcing Arm support for GPU accelerated computing. This may seem strange to some, NVIDIA has been working with Arm for around a decade on the mobile and IoT side. At the same time, developing for GPUs on Arm has been more challenging than x86 or Power because NVIDIA’s tooling was not as mature. At ISC19, NVIDIA announced that they will be changing that, even as they announce a new DGX SuperPOD x86 supercomputer.

NVIDIA Announces Arm Support for GPU Accelerated Computing

At ISC19 NVIDIA says that it will be bringing a full native NVIDIA GPU acceleration platform to the Arm ecosystem. That includes its CUDA-X HPC and AI software stack. The goal is that, by the end of 2019, Arm will be a first-class supported platform like x86 and Power are today.

NVIDIA said that it sees a few trends in the market pushing them towards supporting Arm for HPC and AI:

• The rise in general levels of interest for Marvell ThunderX2 and Ampere’s next-generation parts.
• Japanese and European Exascale projects focused on Arm architectures. Arm is a European country (although based in the UK which is Brexiting the EU) owned by a Japanese firm (Softbank.)
• Strong partner support from companies like Atos, Cray, and HPE (HPE recently announced plans to purchase Cray.)
• A relative open community for Arm server development and the current performance Cavium/ Marvell ThunderX2 parts are seeing.

Overall this makes sense. It is not, however, saying that this is ready. Instead, NVIDIA is saying that this will be ready later this year. There is a big difference, but frankly, by the end of Q3, Arm will need a completely new generation of chips to be competitive on the CPU side of the market.

Final Words

This is a great step for the industry. It also does not necessarily go against the thoughts behind our piece around NVIDIA to Acquire Mellanox a Potential Prelude to Servers. There we speculated that if NVIDIA made a server, it would be Arm based. One would assume that if NVIDIA made an Arm server SoC, it would also need to have CUDA-based HPC and AI tooling working with the Arm ecosystem.

Also, the competitive climate has changed around NVIDIA. With the recent Cray and AMD 1.5 Exaflop Frontier Supercomputer, the first two big exascale contracts have gone to all Intel and all AMD designs, using both companies’ CPUs and GPUs (assuming Intel’s GPU shows up on time and up to spec.) NVIDIA needs to partner with another vendor to get wins. Partnering with IBM Power designs has paid dividends, but we do not see Power as being the top mainstream chip architecture by 2024.