SoftBank has drunk deeply of the AI kool-aid, and now it has its eye on even bigger, smarter large language models (LLMs).
To that end it has deployed what it claims is the world’s largest Nvidia-powered AI supercomputer, to go with the other one it has already built.
Specifics are important whenever declarations like this are made, so here they are: This one is a DGX SuperPOD – the US chip maker’s enterprise-grade AI supercomputer. It is powered by hundreds of its DGX B200 platforms, each one of these being equipped with no fewer than eight Blackwell GPUs.
The end result is a facility boasting 4,000 Blackwells, all connected via Nvidia’s Quantum-2 InfiniBand network architecture – which tops out at 400 Gbps – and running Nvidia’s AI Enterprise software platform.
As for what the Japanese telco will do with its new SuperPOD, why, it will be used for training LLMs, of course.
SoftBank has a dedicated subsidiary for just such a task, called SB Institutions. In fiscal 2024 it trained Japanese-speaking LLMs with a total of 460 billion parameters, and this year it plans to release a commercial model with 70 billion, called Sarashina mini.
SoftBank said its new supercomputer will enable it to accelerate the development of even larger and more advanced models.
It will also serve as a platform for Infrastructure-as-a-Service, with SoftBank making computational resources available to generative AI (GenAI) developers and researchers. In doing so, the supercomputer qualifies for a subsidy under the Japanese Ministry of Economy Trade and Industry (METI)’s Cloud Programme.
Speaking of money, SoftBank didn’t divulge how much it spent on all this, funnily enough. A quick Google reveals that Joe Public can buy a single DGX B200 from Scan.co.uk for a penny less than half a million pounds. Hopefully SoftBank got a bulk discount.
“SoftBank is committed to building different types of social infrastructure to enable a society that co-exists with AI,” the telco said.
“By providing computational resources that can be utilised by a wide range of generative AI developers, SoftBank will contribute to the establishment of development infrastructure and strengthen the service provision framework in Japan.”
Never one for half-measures, SoftBank is going all out to secure its role in the GenAI ecosystem.
It deployed its first DGS SuperPOD, powered by 2,000 Nvidia Ampere GPUs, in September 2023.
In April last year, it was reported that SoftBank planned to spend $960 million over the coming two years to beef up its AI computing facilities.
Indeed, by October 2024, the telco expanded its compute power with the deployment of 4,000 Nvidia Hopper GPUs.
With this new supercomputer, SoftBank now boasts an impressive 10,000 GPUs, which it says is good enough for 13.7 Exaflops – a measure of the performance of a supercomputer that can calculate at least 10¹⁸ floating point operations per second – of total compute power.
As for how and when SoftBank expects to earn a return on all this investment, it depends on whether the LLMs it churns out are actually capable of doing something sufficiently useful that people are prepared to pay for.
While that’s still an open question, research firm STL noted in June that more and more telcos are developing AI projects into commercial propositions. Faint signs of an ROI timeline for SoftBank could soon emerge.
https://www.telecoms.com/ai/softbank-lays-claim-to-world-s-biggest-nvidia-blackwell-ai-supercomputer
