Microsoft Maia: software-defined dataflow and accelerator systems co-design

What changed

Scientific scope

Exposed control/dataflow engines, specialized local memories and synchronization to the software stack, connecting Maia100's hardware/software design to Maia200's inference-focused memory/network hierarchy. Bundles two original technical accounts with separate bylines;2024 is the first cited technical presentation, not a claim of first announcement. SDLA categorization acknowledges existing related architectures rather than inventing all dataflow. Peak FP4/FP8 rates and internal TCO/energy claims are not independent end-to-end certification. Programmer/compiler obligations and workload/network/memory conditions remain; future Maia300 deployment plans are not counted.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Bundles two original technical accounts with separate bylines;2024 is the first cited technical presentation, not a claim of first announcement. SDLA categorization acknowledges existing related architectures rather than inventing all dataflow. Peak FP4/FP8 rates and internal TCO/energy claims are not independent end-to-end certification. Programmer/compiler obligations and workload/network/memory conditions remain; future Maia300 deployment plans are not counted.

Contributors

Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction