Microsoft details Maia 200: no cache hierarchy, software-defined dataflow at 10K TFLOP/s FP4

beenwrekt · x · 2026-08-27

Microsoft published the detailed architecture of Maia 200 (arXiv:2608.24664), its second-gen inference accelerator, already in production in the Azure fleet and built for trillion-parameter frontier models.

The headline decision: no cache hierarchy. The paper argues LLM inference is mostly data-oblivious — compilers can plan nearly every memory access before the model runs, and caches exist to guess what data comes next. When the schedule is known, guessing hardware only adds cost: tag arrays and remapping logic incur 30-35% area and energy overhead plus 10-15% extra access latency.

Maia 200 replaces guessing with orchestration: an architecture called SDLA (Software Defined Locally Accessed Dataflow) exposes every scratchpad, DMA engine, semaphore, and compute unit to software, with C++ control programs running ahead of the data path. Spec-wise, the chip delivers 10,145 TFLOP/s FP4 and 5,072 TFLOP/s FP8 within a 750W TDP, with 7 TB/s HBM bandwidth. This shift from thread-centric to data-movement-centric design yields significant cost and energy savings.

Related event: Microsoft Details Maia 200 Accelerator, Ditches Cache Hierarchy(3 posts)→

Original post →

More from Infra

Infra channel →