Microsoft presented its Maia 200 AI accelerator at Hot Chips 2026, detailing the 3nm chip's architecture ahead of Azure data center deployment. The chip delivers 10,000 TFLOPS of FP4 compute with 7TB/second of HBM bandwidth across six HBM3e stacks, at 750 Watts TDP and an 820mm2 die. Microsoft described its Software Defined Local Access dataflow architecture and scale-up-only Ethernet networking, targeting diverse inference workloads including prefill, decode, and mixture-of-experts models.
No score is assigned. Sources and their independence are shown in the citation chain below.