Microsoft details Azure Maia 300 architecture for agentic inference
Original titleDeep dive into Azure Maia 300 architecture
AISummary
Microsoft describes Azure Maia 300, its third-generation AI accelerator, co-designed across software, silicon, racks, networking, and datacenters.
Compared with Maia 200, the company says Maia 300 delivers up to 7 times greater rack-scale token throughput and more than 2.5 times the performance per watt. The chip adds 288 GB of HBM4 per accelerator and scales to a 12,288-accelerator superpod.
AIWhy it matters
The post gives concrete architecture figures for Maia 300, including compute, memory, network, and rack power, that show how Microsoft targets agentic inference efficiency.
Source: Azure Blog · azure.microsoft.comPublished · added here