Microsoft launches its new Maia 300 AI chip

Microsoft has officially launched its new AI chip called Maia 300. It is built on TSMC's 2nm N2P process and uses a multi-die design with separate memory and I/O dies. Microsoft plans to deploy it in 2027.

For computing power, the chip reaches up to 13.8 PFLOPS in FP4 mode and 6.9 PFLOPS in FP8 mode. It runs at 950W, higher than the 750W of the previous Maia 200 chip.

On the memory side, it packs 288GB of HBM4 with a bandwidth of 18TB/s, making it a strong fit for AI tasks that need lots of memory. It also offers 3.2TB/s of interconnect bandwidth and uses an Ethernet-based network design.

For scale, one rack can hold 96 of these chips, and a cluster can grow to 12,000 accelerators, giving Microsoft more power for large AI workloads.