Microsoft unveils new proprietary AI chip Maia 200

"The accelerator is specifically designed to run large AI models and offers 30 percent better performance per dollar. The chip is manufactured on TSMC's 3-nanometer process and contains more than 140 billion transistors. Microsoft positions Maia 200 as the most powerful first-party silicon of all hyperscalers. In addition to the speed gain on Trainium 3 in FP4 computation (3x faster), Microsoft claims that the chip also outperforms Google's seventh-generation TPU in FP8 computations (5,072 for Microsoft versus 4,614 for Google)."

"Each Maia 200 chip delivers more than 10 petaFLOPS of computing power in 4-bit precision and over 5 petaFLOPS in 8-bit precision. All this within a power budget of 750 watts. The memory subsystem consists of 216GB of HBM3e memory with a bandwidth of 7 TB/s and 272MB of on-chip SRAM. System and network design Microsoft has opted for a two-layer architecture based on standard Ethernet."

"OpenAI uses Maia 200 for GPT-5.2 models, while Microsoft's Superintelligence team uses the hardware for synthetic data generation and reinforcement learning. Maia 200 is currently running in the US Central data center region near Des Moines, Iowa. The US West 3 region near Phoenix, Arizona, will follow soon. Microsoft has not yet announced any other regions in its announcement. Microsoft is now offering a preview of the Maia SDK, complete with PyTorch integration, a Triton compiler, and access to a low-level programming language."

Maia 200 is a Microsoft-designed AI accelerator targeted at running large language and AI models with higher performance per dollar. The chip is built on TSMC's 3-nanometer process and contains over 140 billion transistors. Each chip delivers more than 10 petaFLOPS at 4-bit precision and over 5 petaFLOPS at 8-bit precision within a 750-watt power envelope. The memory subsystem includes 216GB of HBM3e at 7 TB/s and 272MB of on-chip SRAM. Systems use a two-layer Ethernet-based architecture, offer 2.8 TB/s bidirectional system bandwidth, and can scale to clusters of up to 6,144 accelerators. Maia 200 is integrated into Azure services and available with an SDK, simulator, and cost tools for developers.

Read at Techzine Global