Microsoft details Azure Maia 300 AI accelerator architecture; third-generation design delivers up to 7x rack-scale token throughput vs Maia 200
Microsoft published a technical deep dive on its third-generation Azure Maia 300 AI accelerator, describing the system-on-chip, rack-scale, and superpod design intended for cloud-scale agentic inference. Relative to Maia 200, Microsoft states Maia 300 delivers up to 7x greater rack-scale token throughput in a 96-accelerator configuration and more than 2.5x the performance per watt. The article does not disclose revenue, profit, or other financial results and does not provide financial guidance.
Key takeaways
Maia 300 is Microsoft's third-generation Azure AI accelerator, optimized for long-context, reasoning-intensive, and decode-heavy LLM inference.
Compute dies are fabricated on TSMC N2P; memory/I/O functions are placed on separate chiplets, with a NUMA-style organization across chiplets.
Each accelerator includes 8 HBM4 stacks providing 288 GB of capacity and 18 TB/s of bandwidth.
System compute FLOPs are stated to rise 36%, HBM bandwidth 2.5x, and scale-up bandwidth 14% versus Maia 200.
Rack scale is 96 accelerators across 24 four-accelerator trays within a 167-kW rack, doubling accelerator density versus Maia 200.
A two-tier Ethernet-based scale-up fabric supports superpods of up to 12,288 accelerators with a single communication model.
Compared with Maia 200, Microsoft states Maia 300 delivers up to 7x rack-scale token throughput and more than 2.5x performance per watt.
Key metrics
Primary source
TickerStar generated this brief from an official company, investor-relations, earnings-call, or regulatory source. Review the primary material before making an investment decision.
Open Investor relationsFollow MSFT in real time
Get material earnings, filing, IR, and earnings-call alerts by email.