Deep dive into Azure Maia 300 architecture
Summary
Azure Maia 300 is a next-generation AI accelerator system designed to optimize LLM inference with a software-defined dataflow (SDLA) architecture. It evolves Maia 200 by introducing a locality-aware non-uniform memory access (NUMA) layout across chiplets, narrower precision formats, and a large-scale network fabric to support a scale-up design. The rack supports up to 12,288 accelerators in a superpod, delivering higher tokens-per-watt for cloud AI workloads.