From the source
Microsoft published a deep dive into the architecture of its third-generation AI accelerator, Maia 300, which is designed for cloud-scale agentic inference.
The system-on-chip uses a multi-chiplet design on TSMC N2P, introduces enhanced MXFP4 and sub-3-bit storage formats, and provides 2.5 times the memory bandwidth and 33% more capacity than Maia 200.
The rack doubles to 96 accelerators and the scale-up domain grows to a superpod of 12,288 accelerators.
