Like many of his peers, Microsoft CEO Satya Nadella wants to mark a new golden age of its technology thanks to AI. However, he is not convinced that the fate of this dream can be beyond his control.
That’s why Nadella tasked his team with building Microsoft its own powerful silicon chip to disrupt a market that, until now, Nvidia is largely blocked.
After watching India face New Zealand in Cricket World CupNadella took the stage at Microsoft’s Ignite conference to reveal the tech giant’s first-ever custom silicon called the Azure Maia 100.
The in-house designed AI “accelerator” – a fancy name for a type of semiconductor – is designed for use in ultra high bandwidth data centers that power large language models (LLM).
Maia is a potential alternative to the Nvidia A100 and H100 processors that currently dominate the AI chip market.
“We are already testing this with several of our own AI services, including the GitHub Copilot,” Nadella said. Wednesday. “We will deploy Maia accelerators across our entire fleet, first supporting our own workloads, and then scaling them to third-party workloads. »
One such potential customer is Sam Altman of OpenAI. The CEO of ChatGPT considers Microsoft a major shareholder and, until now, he has had to rely on the more and more rare silicon from Nvidia to power its AI chatbot.
“We were excited when Microsoft first shared its designs for the Maia chip, and we worked together to refine and test it with our models,” Altman said in a statement. statement Wednesday.
Nadella said the workforce will enter a “co-pilot era” driven by AI, defined by cutting-edge generative artificial intelligences with supercomputer performance with human-like behavior.
The Maia chip includes 105 billion transistors using 5-nanometer nodes, the most advanced commercially available technology for miniaturizing semiconductor circuits.
It also comes with its own “companion” thermal management system that replaces conventional fans with sophisticated liquid cooling for superior performance.
“The power demands of AI require a radically different infrastructure than other clouds,” Nadella explained.
Along with their sidekick, the chips are then mounted on specially designed server racks that can be dropped into existing Microsoft data centers rather than requiring new investment.
Nvidia struggles to meet demand from buyers like Elon Musk
The Azure Maia 100 chip represents the final piece of the puzzle for Microsoft to deliver a complete system capable of handling the heavy workloads needed for AI training and inference.
By controlling every facet, including the required software, server rack, and cooling system, custom design can better ensure the whole is greater than the sum of its parts.
“Azure’s end-to-end AI architecture, now optimized down to silicon with Maia, paves the way for training better models and making those models less expensive for our customers,” said Altman of OpenAI.
He joined Nadella on stage at the Ignite conference to talk about their continued collaboration across various areas of the tech industry.
After all, Microsoft isn’t switching to Maia en masse. Nadella committed to offering Huang’s latest AI training chip, the H200 tensor core GPU revealed this weekto Microsoft data center customers starting next year as an alternative.
“It’s hard to believe it’s only been a year since ChatGPT was released,” Nadella said. “We are at a critical point. This is clearly the era of co-pilots.