Overview
- Microsoft announced the Maia 300 as its third‑generation in‑house AI accelerator built to run large models for Azure and its internal AI services.
- The Maia 300 SoC pairs 288GB of HBM4 memory with 18TB/s of memory bandwidth and delivers peak FP4 performance of 13.8 PFLOPS and FP8 of 6.9 PFLOPS.
- Microsoft designed Maia 300 as a multi‑chiplet package that uses TSMC N2P (2nm) for the main compute chiplets and separate chiplets for memory and I/O.
- The company published system figures showing 96 chips per cabinet with an estimated cabinet power around 167 kW and a per‑chip rated power of 950W, and it plans deployments at more than one million chips and multi‑gigawatt scale starting in 2027.
- Microsoft says full‑stack optimization gives Maia 300 roughly 40% more effective compute per megawatt than mainstream GPU systems, but those efficiency and cost claims have not been independently validated and will depend on real‑world deployments, power and cooling tradeoffs, and foundry supply.