Overview
- Built on TSMC’s 3nm process, Maia 200 integrates 216 GB of HBM3E via advanced packaging for memory‑intensive inference workloads.
- Microsoft claims FP4 throughput triples Amazon’s Trainium v3, FP8 performance exceeds Google’s TPU v7, and performance per dollar is about 30% higher than its current data‑center hardware.
- The accelerator is live in Microsoft’s Des Moines, Iowa region with a planned rollout to the Phoenix, Arizona US West 3 region.
- Microsoft promotes a more efficient water‑cooling design with a stated goal of achieving “zero waste” to reduce local environmental and water pressures.
- Korean media report SK Hynix as the exclusive HBM supplier using six HBM3E stacks per chip, and the stock rose 8.7% after the reports; SK Hynix declined to comment.