AI platform now has 6,000 GPUs installed
SoftBank Corp has finished installing 4,000 Nvidia H100 GPUs in its Japan-based artificial intelligence (AI) computing platform.
The computing platform originally began operations in September 2023, then with 2,000 Nvidia Ampere GPUs and 700 petaflops of AI computing performance, which is typically measured in FP8 or eight-bit precision calculations rather than FP64 double-precision calculations, also known as FP64.
The latest installation brings its total to 6,000 GPUs and a performance of approximately 4.7 exaflops of AI compute.
The company has further plans to expand this with an Nvidia DGX SuperPOD featuring GGX B200 systems within the fiscal year (FY) 2025, ultimately growing the number of GPUs installed to around 10,000 with a compute capability of 25.7 exaflops.
This will initially be used by SoftBank’s subsidiary dedicated to developing Japanese large language models (LLMs), SB Intuitions, which is aiming to develop LLMs with 390 billion parameters that can support multiple modalities, within FY2024.
SoftBank will also provide access to the platform as an “Infrastructure-as-a-Service” offering to companies and research institutions in Japan.
Where the AI computing platform is hosted is not specified. SoftBank operates 13 data centers across Japan; four in Tokyo, three each around Osaka and Kitakyushu, and one each in Oita, Fukushima, and Sapporo.
Earlier this year the company revealed plans to develop a 150MW data center at a soon-to-be-closed liquid crystal display (LCD) panel-making site in Osaka.
SoftBank is a Japanese multinational investment holding company that primarily invests in companies operating in technology sectors.
Earlier this year, SoftBank was reportedly in talks with TSMC to develop its own AI chips. The company’s founder Masayoshi Son was reportedly looking to raise $100 billion for an AI chip venture dubbed Izanagi to compete with Nvidia first surfacing in February 2024.
In July, SoftBank acquired struggling British AI chip designer Graphcore for an unconfirmed amount. The conglomerate also owns a majority stake in Arm, having bought the British chip designer for $32 billion in 2016 and retained a controlling interest in the company when it re-floated on the stock market last year.
https://www.datacenterdynamics.com/en/news/softbank-installs-4000-nvidia-h100s-for-japanese-ai-computing-platform/
