HPE is to build a supercomputer for Japan’s AIST research institution, using thousands of Nvidia’s latest H200 GPUs to support large foundational models for generative AI in research.
The ABCI 3.0 (AI Bridging Cloud Infrastructure) system is to be offered as a cloud service to both public and private organizations by AIST (National Institute of Advanced Industrial Science and Technology). Its purpose is ultimately the advancement of research, development, and innovation, HPE said.
Or if you are Nvidia, it is “enhancing Japan’s AI sovereignty” as well as strengthening its research and development capabilities.
HPE is building the system using Cray XD boxes interconnected by Nvidia’s Quantum-2 InfiniBand networking. This is likely the Cray XD670, which can fit 8 x Nvidia H200 or H100 GPUs into a 5U rack-mount chassis and is powered by Intel’s “Emerald Rapids” 5th Gen Xeon Scalable processors.
The Texas-based IT titan said that ABCI 3.0 is expected to be the fastest AI supercomputer in Japan when it comes online, and claims it will achieve a theoretical maximum of approximately 6.2 exaflops of half-precision (16-bit) performance. But AI exaflops figures should always be taken with a pinch of salt, as The Register has explained previously.
According to Nvidia, the system will offer 410 double-precision petaflops, a measure of general computing capacity.
Details of the total number of nodes, the cost of the system and when it will be operational were not announced. We asked HPE, but it said it was not able to disclose that information at this time.
AIST said it has made its supercomputing services accessible via the cloud to research institutions, universities, and commercial companies of all sizes since the first ABCI in 2018, and these have so far contributed to the development of AI techniques to predict genome sequencing as well as Japanese large language models (LLMs).
“In response to the escalating demand for generative AI and computational resources, AIST is maximizing the capabilities of ABCI 3.0 with unparalleled machine learning performance delivered through a powerful supercomputer,” AIST Executive Officer Dr Yoshio Tanaka said in a statement.
The ABCI 3.0 project is being built with support from Japan’s Ministry of Economy, Trade and Industry (METI) to strengthen the nation’s computing resources via the Economic Security Fund. It is part of a broader $1 billion initiative by METI that includes both ABCI efforts and investments in cloud AI computing.
Nvidia’s mention of “AI sovereignty” is a recent theme of CEO Jensen Huang, who says that sovereign AI refers to a nation’s ability to harness artificial intelligence via its own compute infrastructure and resources, rather than relying on others.
“Every nation needs its own AI infrastructure to drive innovation and growth,” Huang said.
Kernel tweaks improve Raspberry Pi performance, efficiency
Musk ropes Dell, Supermicro into xAI supercomputer project
ASUS quietly built supercomputers, datacenters and an LLM. Now it’s quietly selling them all together
Memphis to host ‘Gigafactory of Compute’ thanks to xAI and Elon Musk’s billions
Another hobby horse of his is “AI factories” – datacenters specially built to handle the most computationally intensive AI tasks.
“The AI factory will become the bedrock of modern economies across the world,” Huang told the Japanese press during a visit to the country.
Perhaps coincidentally, Europe’s supercomputing body the EuroHPC JU announced this week that an “AI Factories” amendment to its official regulations had come into force. This means it is now encouraged to acquire and operate dedicated AI-optimized supercomputers, with the aim of developing AI ecosystems centered on EuroHPC supercomputing facilities. ®
>>> Read full article>>>
Copyright for syndicated content belongs to the linked Source : The Register – https://go.theregister.com/feed/www.theregister.com/2024/07/12/hpe_to_build_supercomputer_to/