跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators未公布单价No public price

Microsoft Maia 200(推理)Microsoft Maia 200 (inference)

未公布单价No public price未公布单价No public price · 官方页未给公开单价;这类产品通常按整机、云实例或询价成交,不拿经销商价冒充官方价。No public unit price on the official page; these usually sell as systems, cloud instances or on request, so dealer quotes are not presented as official.

回答「Microsoft 自研推理芯片到哪一代、规格和部署状态如何」。Maia 200 是微软第二代自研 AI 加速器,TSMC 3nm、216GB HBM3e / 7TB/s、272MB 片上 SRAM、原生 FP8/FP4,单芯片 FP4 超过 10 PFLOPS、FP8 超过 5 PFLOPS,750W SoC;2.8TB/s 双向 scale-up,官方称可连 6,144 颗。已部署在 US Central,US West 3 随后。
不主张:Maia 200 不对外以 Azure VM 形式出租,Azure 客户是通过 Microsoft 365 Copilot、Foundry 等服务间接用到;30% 每美元性能是微软对自家 fleet 的口径,不是跨厂商横评;没有公开价格;10 PF 是 FP4,不是所有精度;6,144 是 scale-up 集群上限,不是单机架。
Answers which generation Microsoft's in-house inference chip has reached, its specs and deployment state. Maia 200 is Microsoft's second-generation custom AI accelerator: TSMC 3nm, 216 GB HBM3e at 7 TB/s, 272 MB on-chip SRAM, native FP8/FP4, over 10 PFLOPS FP4 and over 5 PFLOPS FP8 within a 750 W SoC, plus 2.8 TB/s bidirectional scale-up and clusters of up to 6,144 accelerators. It is deployed in US Central, with US West 3 next.
It does not claim: Maia 200 is not rented as an Azure VM; Azure customers reach it through services such as Microsoft 365 Copilot and Foundry. The 30% performance-per-dollar figure is Microsoft's comparison against its own fleet, not a cross-vendor benchmark; no public price; 10 PF is FP4, not every precision; 6,144 is a scale-up cluster ceiling, not a single rack.

关键参数Key specs

定位 / role
大模型推理
制程 / process
TSMC 3nm
晶体管 / transistors
超过 1400 亿
显存 / memory
216 GB HBM3e
带宽 / bandwidth
7 TB/s
片上 SRAM / SRAM
272 MB
FP4 / FP8
>10 PFLOPS / >5 PFLOPS
功耗 / TDP
750 W SoC
Scale-up
2.8 TB/s 双向;最多 6,144 颗
部署 / deployment
US Central 已部署;US West 3 随后
状态 / status
2026-01-26 发布

官方原文怎么说What the source says

an accelerator built on TSMC’s 3nm process with native FP8/FP4 tensor cores, a redesigned memory system with 216GB HBM3e at 7 TB/s and 272MB of on-chip SRAM
1 https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
each Maia 200 chip delivering over 10 petaFLOPS in 4-bit precision (FP4) and over 5 petaFLOPS of 8-bit (FP8) performance, all within a 750W SoC TDP envelope.
2 https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
2.8 TB/s of bidirectional, dedicated scaleup bandwidth
3 https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
Predictable, high-performance collective operations across clusters of up to 6,144 accelerators
4 https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
30% better performance per dollar than the latest generation hardware in our fleet today
5 https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/

资料与许可Source and licence

资料时间Source checked
2026-11-20
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
类目Category
加速卡Accelerators
状态Status
已发布 / 未出货Announced
适合Use cases
训练Training
价格口径Price basis
未公布单价No public price · not-published

同类目还有More in this category

下一步Next step

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute