AI Inference Is a Data Movement Problem: Microsoft's Saurabh Dighe on Maia 300

Microsoft's Maia 300 is a fully co-designed system, optimized to solve the memory-bound "data movement problem" of emerging agentic AI workloads.

Saurabh Dighe, Corporate VP of Azure Hardware Systems at Microsoft, sits down with Austin to detail the architecture, from its TSMC N2P chiplet design enabling 18 TB/s of HBM4 memory bandwidth to its use of AI agents for automated kernel optimization.

They discuss how this full-stack approach aims to maximize intelligence per watt and per dollar for Microsoft's planet-scale services.

This episode is presented by @CrusoeAI. Crusoe's Serverless Fine-Tuning trains an open model like Llama, DeepSeek, or GLM on your own data, and you pay per million tokens, not per GPU hour. About $2 buys roughly 5M training tokens on smaller models. Get $5 in credit to start: https://semidoped.com/crusoe

Key Takeaways:

- Microsoft views AI inference as a data movement problem, not a compute problem. The energy cost of moving data to feed the compute engine can be two orders of magnitude higher than the compute itself.

- The design of Maia 300 is driven by the shift from chat to agentic AI, which creates memory-bound workloads. The system delivers a 7x token throughput increase per rack over Maia 200, with a 5x increase in rack-level memory bandwidth.

- Maia 300 uses a chiplet architecture to "extend the shoreline" of its single TSMC N2P compute die. This allows it to accommodate eight HBM4 stacks, delivering 18 TB/s of memory bandwidth.

- Microsoft is using AI agents for "agentic kernel optimization" to automate the creation of low-level software. This strategy aims to bypass the manual optimization bottleneck and enable "Day-zero model availability" for new AI models.

- The system is built for deployment flexibility. While servers use direct liquid cooling, a rack-level "heat exchanger unit" allows the 167 kW racks to be deployed in traditional air-cooled data centers.

- The network is a two-tiered scale-up fabric built on standard Ethernet. Copper connects the 96 accelerators within a rack, while optics connect racks into superpods of up to 12,288 accelerators.

Chapters:
0:00 Intro: Saurabh Dighe
1:39 Microsoft's Silicon Philosophy
5:52 Workload Shift: Agentic AI
9:32 The Data Movement Problem
14:20 Crusoe Ad Read
17:47 Maia 300 Specs
24:07 Chiplets & HBM4
29:12 Prefill/Decode Disaggregation
30:33 Two-Tiered Network
35:45 NPO vs. CPO
36:46 Rack-Scale Power & Cooling
41:35 Agentic Kernel Optimization
49:08 What's Next: Maia 400

Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat

Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr

Follow Semi Doped:
Get more of Austin and Vik daily, free: https://daily.semidoped.com/