Microsoft launches Maia 200 AI accelerator for Azure Cloud Infrastructure

Maia 200 is specifically designed for compute?intensive AI inference and integrates seamlessly into Microsoft Azure.

Microsoft has announced Maia 200, an AI accelerator designed for inference workloads in its Azure cloud platform. The chip is built for deploying AI models in production applications.

The accelerator's architecture includes compute components, a memory system, and networking capabilities. Microsoft plans to use Maia 200 across its cloud infrastructure for AI workloads. The chip is designed for compute-intensive AI inference tasks and operates within Microsoft Azure.

Technical Specifications

Maia 200 provides over 10 petaFLOPS at 4-bit precision (FP4) and more than 5 petaFLOPS at 8-bit precision (FP8). The chip is manufactured using 3-nanometer process technology. The networking architecture supports clusters of up to 6,144 AI accelerators over standard Ethernet.

The chip can run current large-scale AI models and is designed to support future model sizes.

Deployment and Applications

Microsoft will initially deploy Maia 200 in U.S. regions of Azure. The accelerator will run AI models from Microsoft's Superintelligence team. Applications include Azure AI Foundry, Microsoft's platform for developing AI applications and agents, and Microsoft 365 Copilot.

Microsoft designs chips, develops AI models, and builds applications within a single operational framework. The company runs AI workloads that inform chip design, model development, and application optimization.

Software Development Kit

Microsoft is releasing a preview of the Maia Software Development Kit (SDK). The SDK supports AI frameworks and enables developers to optimize models for Maia systems. The SDK includes a Triton compiler, PyTorch support, NPL programming, a simulator, and a cost calculator.