Monitor this situation.
Unsubscribe anytime.
[SITUATION] · [QUIET] · [TECHNOLOGY]
2 clusters · 5 sources · 4 days · First seen · Last updated
Microsoft Maia 300 AI accelerator development
Overview
Microsoft has unveiled the Maia 300, the third generation of its proprietary AI accelerator platform. Designed primarily for inference tasks rather than training, the hardware is intended to handle memory-intensive models and complex logical reasoning within Microsoft products.
The platform utilizes a modular chiplet architecture manufactured by TSMC using a 2-nanometer class process. Each accelerator features 288 GB of HBM4 memory with a bandwidth of 18 TB/s. The system is designed for massive scale, supporting up to 96 accelerators per rack and the ability to interconnect up to 12,288 accelerators into a Superpod via a custom Ethernet-based AI network fabric.
Microsoft expects production to begin in 2027. The company claims the Maia 300 will provide approximately 40 percent more tokens processed per megawatt compared to leading GPU systems. However, some industry experts have questioned this metric, arguing that “tokens per watt” is an ambiguous measurement that lacks a standardized definition.
Entities
Microsoft · TSMC · Maia 300 · Scott Guthrie · Dell Technologies
Timeline
-
[TECHNOLOGY] 3 sourcesMicrosoft announces Maia 300 AI accelerator
Microsoft announced the Maia 300 AI accelerator, a third-generation chip designed for high-efficiency inference, featuring 2-nanometer architecture and significant throughput improvements.
-
[TECHNOLOGY] 2 sourcesMicrosoft unveils Maia 300 AI accelerator platform
Microsoft revealed its Maia 300 AI accelerator, a high-bandwidth, liquid-cooled platform optimized for inference at massive scale, despite industry debate over its efficiency metrics.
Sources
ad-hoc-news.de · drweb.de · infoq.cn · it-daily.net · servethehome.com