Microsoft today unveiled the Microsoft Azure Maia 300, its largest in-house compute processor for Azure Cloud deployment. CEO Satya Nadella announced plans to scale the Maia chip family to multi-gigawatt deployments, positioning it as the company's primary AI inferencing system.
The Maia 300 builds on the Maia 200, currently deployed across US, Europe, and Australia regions, which delivers 70% more tokens per Watt than AI GPUs. The new processor offers a further 40% uplift in tokens per Watt efficiency.
The multi-chiplet tile-based processor features a main compute die built on TSMC's N2P (2 nm) node, 18 TB/s HBM4 bandwidth, and supports up to 96 chips per rack (Maia Pod) with scale-out to 12,000 chips within a domain. It delivers 13.8 peta-OPS FP4 and 6.9 peta-OPS FP8 performance with a 950W TDP.
Source: TechPowerUp News · Summarized by HeadlinesBriefing