重點資訊
Microsoft has officially launched the Azure Maia 300, its most powerful in-house compute processor designed specifically for AI inferencing within its cloud infrastructure. CEO Satya Nadella confirmed plans to scale the Maia chip family into multi-gigawatt deployments to support the company's internal products and services.
The Maia 300 utilizes a multi-chiplet tile architecture featuring a main compute die manufactured on the TSMC N2P 2 nm node. It incorporates 288 GB of HBM4 memory across eight stacks, delivering 18 TB/s of bandwidth. The processor has a maximum thermal design power of 950 W.
Performance metrics for the new silicon include 13.8 peta-OPS for FP4 throughput and 6.9 POPS for FP8. The chip supports a 3.2 TB/s inter-chip link and is compatible with a two-tiered AI Transport Ethernet network interface, allowing for massive scaling up to 12,000 chips within a single domain.
This release follows the deployment of the Maia 200, which currently operates in the US, Europe, and Australia. Microsoft claims the Maia 300 will provide a 40% improvement in tokens per Watt compared to existing GPU-based solutions, building on the 70% efficiency gain already observed with the Maia 200.
