What Microsoft, Nvidia AI supercomputer partnership means

Nvidia and Microsoft now have a multi-year partnership to build a cloud AI supercomputer that will enable companies to build large language models and complex recommendation systems.

The vendors revealed on November 16 that the supercomputer will be powered by Microsoft Azure’s infrastructure and Nvidia’s GPUs, networking and AI software stack.

Fundamental models and generative AI

Nvidia will also use Azure’s VM instances for research and development in generative AI and basic AI models. Generative AI and basic models such as GPT-3 enable enterprises and organizations to create text, images or code. Meanwhile, Microsoft customers can access Nvidia’s full stack of AI workflows and software kits.

Azure’s AI supercomputer will be the first public cloud supercomputer to include Nvidia’s AI stack with A100 and H100 GPUs, the Nvidia Quantum-2 400 Gbps InfiniBand network system and the Nvidia AI Enterprise software suite, the vendors said.

The collaboration comes as generative AI and base models become increasingly popular among organizations since the release of GPT-3, image generation tools such as Dall-E and Stable Diffusion, and Microsoft’s GitHub Copilot.

Microsoft’s search for alternatives

The partnership shows that Microsoft has increased demand from its customers for access to Nvidia GPUs, particularly H100s, said Karl Freund, founder and analyst at Cambrian AI Research.

You want to make it big in supercomputing, so with this announcement they could say, “We’re doing something big here.”

Charles FreundFounder and Analyst, Cambrian AI Research

“It’s a little surprising that they came out with this big announcement for supercomputing, but not a big deal,” Freund said. “You want to be big in supercomputing, so with this announcement they could say, ‘We’re doing something big here.'”

While Microsoft has had a relationship with Nvidia in the past, it has also shown interest in alternatives to the chip supplier. For example, Microsoft entered into a partnership with Graphcore in 2019. The chip startup had a deal to supply Microsoft with processors for its cloud computing platform. However, that collaboration has since ended.

“Nobody can even come close to Nvidia’s software,” Freund said, adding that while competitors like Intel or Graphcore can compete on performance, they can’t compete effectively in the software arena. “If you’re Microsoft…there’s really only one choice, and that’s Nvidia.”

If Nvidia can deliver on H100s, the partnership should go well, he said. This means enterprises should see thousands of Nvidia GPUs on Azure by the first quarter of next year.

The alliance also provides insight into what infrastructure Microsoft and OpenAI — in which Microsoft has invested $1 billion — could use to develop GPT-4, the next generation of the powerful large language model, he continued.

“My assumption is it’s Nvidia A100,” Freund said. “If … it’s not Nvidia A100, then Microsoft probably won’t be able to announce this major commitment to continue investing in Nvidia hardware.”

Meanwhile, both Nvidia and Microsoft stand ready to support enterprises as they increasingly train generative AI models as well as base models.

“These giant base models and generative AI models aren’t just trained on small-scale infrastructure,” said Paresh Kharya, senior director of product management and marketing for accelerated computing at Nvidia. “It’s the full stack that’s really important to be able to offer really attractive and efficient bidding [products] to meet business demand.”

Leave a Comment