Microsoft is reviving its Maia custom AI chip program after facing delays and staffing challenges. The Maia 200 inference chip, built on TSMC’s 3nm process, is already being deployed in Iowa and Arizona data centers, featuring 216GB of HBM3e memory and native support for FP8 and FP4 precision formats. It delivers three times the FP4 performance of Amazon’s Trainium3 and roughly 30 percent better performance-per-dollar compared to the previous generation. It is already powering Microsoft 365 Copilot and OpenAI models. The successor Maia 300 is in design, with development timeline extending into 2027.
Source: Read the original article
