微软宣布自明年起将 Maia 芯片扩展到多吉瓦部署，作为自家产品与服务的主力推理系统，并发布 Maia 300（每瓦 token 数比领先 GPU 系统高 40%）。
AI 生成摘要 · 以原文为准
A few more interesting nuggets on this announcement as $MSFT starts to move custom silicon into their scale up compute domain.
- A Maya 300 rack is planned to contain 96 accelerators at approximately 167 kilowatts and can use liquid-cooled chips with either liquid-cooled facilities or heat-exchanger units in air-cooled data centers.
- Microsoft projects up to seven times more tokens per rack than Maya 200, 1.4 times more tokens per gigawatt, and 30% better tokens per dollar.
- The architecture uses Ethernet-based scale-up networking, custom transport, DMA, and two-tier switching to connect across racks and potentially colos. Current deployments are centered on 64- to 128-accelerator clusters, with 12,000 accelerators representing a current plan of record.
Today we are announcing our biggest bet yet on our own silicon. Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services. And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems. We are already deploying Maia 200 across the US, Europe, and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot. Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet.在 X 查看被引用的帖子
来源：Ben Bajarin · x.com