Microsoft Unveils Maia 200 AI Chip for Scaled Inference

[chainsniff]

Microsoft has announced the launch of its new Maia 200 chip, a custom-designed processor built to scale artificial intelligence inference. The Maia 200 follows the Maia 100 released in 2023 and is positioned as a major performance upgrade aimed at running large AI models faster and more efficiently.

According to Microsoft, the chip contains more than 100 billion transistors and delivers over 10 petaflops of 4-bit performance, along with roughly 5 petaflops at 8-bit precision. The company says this allows a single Maia 200 node to run today’s largest models with room to support even more demanding workloads in the future.

The chip is designed to reduce power use and operational disruption as AI systems scale.

[ed]

finally a chip actually built for inference not just flexing on training numbers