IT之家 News on October 12: Microsoft today officially released the Maia 300 AI chip. This product is equipped with 288GB of HBM4 memory, with bandwidth reaching 18TB/s, FP4 compute power of up to 13.8 PFLOPS, built on TSMC's N2P process, and is planned for deployment in 2027.
IT之家 noted that the Maia 300 adopts a multi-chiplet design,with the main compute portion built on TSMC's N2P 2nm process, integrating independent memory chiplets and I/O chiplets.FP4 compute power reaches up to 13.8 PFLOPS, and FP8 compute power is 6.9 PFLOPS. Its SoC power rating is 950W, significantly higher than the Maia 200's 750W.
In addition, the product is equipped with 8 stacks of HBM4 memory,with a total capacity of up to 288GB, and bandwidth of 18TB/s, making it better suited for AI inference tasks with high memory requirements. Bidirectional scale-out interconnect bandwidth reaches 3.2TB/s, using a two-layer Ethernet-based network architecture. A single rack can hold 96 Maia 300 chips, and clusters of up to 12000 accelerators can be built.