Skip to content

Maia 100

A custom AI accelerator built for Azure workloads

Visit official site
Category AI models

What Maia 100 is

The chip is described as Microsoft's first custom AI accelerator designed for Azure. Hardware and software are co-designed for large-scale AI workloads with high bandwidth and capacity.

What you can do with it

  • Run AI workloads on Azure with 1.8TB/s bandwidth
  • Use tensor units and a vector processor for mixed data types
  • Connect chips over Ethernet for collective traffic

Application scenarios

  • Train and infer deep learning models on Azure
  • Move PyTorch or Triton models with the Maia SDK
  • Tune custom kernels for a specific workload

Getting started

  • Create a Maia 100 instance on Azure
  • Migrate a model with the Maia SDK
  • Compile the model into an executable format

Pros & cons

✓ What we like

  • Ethernet interconnect supports large-scale model traffic
  • The Maia SDK deploys PyTorch and Triton models with two programming models

! What to watch out for

  • The chip is tied to the Azure platform
  • Deployment goes through the Maia SDK toolchain

FAQ

How much memory is stated?

The material gives 1.8TB/s bandwidth and 64GB of capacity.

What does the SDK support?

The feature list names fast deployment of PyTorch and Triton models and a dual programming model.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review