Microsoft recently showcased the third-generation of its custom in-house AI chip, as part of the organisation’s plan to reduce reliance on purchasing processors from Nvidia.
The Maia 300 chip will be used for the tech giant’s multi-gigawatt deployments in 2027, making it the primary inference system for Microsoft products and services. Reuters reported it plans to potentially deploy more than 1 million of the chips next year.
According to a blog post from EVP Scott Guthrie, the company expects Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems without naming Nvidia.
The Maia 300 is built as a single system, with software, silicon, physical rack infrastructure, networking, and data centres co-designed to operate as a single platform.
Guthrie noted a 96-chip rack offers seven times the rack-scale token throughput of Microsoft’s Maia 200. He said the new chip is “designed with an inference-first architecture optimised for long-context, memory-intensive and decode-heavy AI workloads”.
The Maia 200 chip powers the company’s production workloads, including Microsoft 365 Copilot, as well as OpenAI models running on Microsoft’s Azure cloud computing service.
Source: Mobile World Live
Image Credit: Microsoft
Source: Tahawul Tech