Microsoft Publishes Maia 200 Specs: 10,145 Tflop/s FP4 and 7 TB/s HBM in a 750W Envelope

arXiv 2608.24664·medium signalXBlueskyLinkedInCopy link

Maia 200 is presented as an AI accelerator delivering 10,145 Tflop/s FP4 and 5,072 Tflop/s FP8 within a 750W TDP alongside 7 TB/s HBM bandwidth. The architectural claim is a category the authors call Software Defined Locally Accessed Dataflow Architectures, which explicitly program dataflow engines to orchestrate specialized memories and data movement engines, shifting from thread-centric to data-movement-centric design. They accompany it with a taxonomy of data management inspired by Flynn's classification, and target inference workloads specifically on cost and energy grounds. Treat the throughput figures as vendor-reported until third-party benchmarks appear.