"Nvidia brings CUDA to Arm," the graphics chip giant announced with a straight face on Monday, specifically, "its support for Arm CPUs."
What's curiously omitted from today's marketing blurb, though, is any concrete reference to Nvidia's existing support for CUDA and, er, Arm.
Anyone who has used one of, say, Nvidia's automotive and internet-of-things development boards, which sport system-on-chips that feature a mix of Arm Cortex CPU cores and Nv CUDA-based GPU engines, will know Nvidia already develops and distributes Arm-compatible CUDA libraries and toolkits. Arm support is already there for CUDA.
For the uninitiated, CUDA is a programming interface for turning Nvidia graphics processors into math accelerators for neural network applications, simulations, and other vector-intensive software.
What appears to have been announced this week – in time for the International Supercomputing Conference (ISC) in Frankfurt, Germany – is that Nvidia will, by 2020, expand its Arm support for CUDA graphics chips, and go beyond the usual embedded electronics, the Internet of Things, and automotive projects, to include high-end GPU accelerators found in supercomputers, beefy servers, and other big iron.
This will, when it arrives, put Arm CUDA support on the same footing as that of IBM POWER, and Intel and AMD x86 processors, we're told. Will RISC-V be next, given that Nvidia is a founding platinum member of the open-source ISA's foundation?
According to the blurb:
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.
Once stack optimization is complete, Nvidia will accelerate all major CPU architectures, including x86, POWER and Arm.
The key words here are "full stack," we guess. We asked an Nvidia for further comment on how today's announced Arm support compares to Nvidia's previous Arm CUDA support. "We did not support CUDA-X libraries, and all our dev tools like compilers, on Arm before this announcement," a spokesperson told us.
In a canned statement, Jensen Huang, founder and CEO of Nvidia, proclaimed: "As traditional compute scaling ends, power will limit all supercomputers. The combination of Nvidia's CUDA-accelerated computing and Arm's energy-efficient CPU architecture will give the HPC community a boost to exascale."
HPC processor project tosses architectural designs on desk of the European CommissionREAD MORE
The firm boasts its GPUs are included in 22 of the world's 25 most energy-efficient supercomputers, as noted on the Green 500 list.
The great and good of the exascale world stepped up to praise Nvidia. Arm CEO Simon Segars said the Nvidia collaboration "is a key milestone for the HPC community."
Cray CEO and president Peter Ungaro said that his company would be integrating Nvidia CUDA and CUDA-X HPC and AI software stack support "closely with our Cray system management and programming environment (compilers, libraries and tools) already enabled to support Arm processors across our XC and future Shasta supercomputers."
EPI general manager Philippe Notton chimed in: "The combination between the EPI Arm-based microprocessor and Nvidia accelerator could make a perfect match for equipping building blocks in the future European exascale modular supercomputers."
HPE's Bill Mannel, VP and GM of HPC and AI, said: "Nvidia's support for Arm complements our latest developments on the HPE Apollo 70, an Arm-based, purpose-built HPC system, and now, Nvidia GPU-enabled."
Japan's Post-K exascale supercomputer, being built by Fujitsu, will use homegrown 64-bit Arm processors, as will a proposed European EPI exaflopper, while other exascale systems are set to employ Intel and IBM processors. Nvidia's expanded Arm CUDA support comes as there is growing interest in Arm and RISC-V-based supers that rely on graphics engines and/or custom accelerators. ®
Nvidia's other announcements at ISC '19...
- A 9.4-petaFLOPS DGX SuperPOD to provide AI infrastructure for the training of self-driving vehicles. It uses 96 x DGX-2H systems, with 1,536 Nvidia V100 Tensor Core GPUs, plus Mellanox interconnect tech. Mellanox is being bought by Nvidia for $6.9bn.
- Nvidia's NGC unit offers more than 50 GPU-optimised containers for deep learning frameworks, machine learning algorithms and HPC applications that run on Docker and Singularity. Japan's ABCI supercomputer, with Nvidia Tensor Core GPUs, has adopted these containers.
- CUDA-X HPC is similar to CUDA-X AI, and is a set of tuned kernels, libraries, tools, compilers and APIs for HPC use. Find out more here.