从 2027 年开始，微软将把 Azure Maia 系列的部署规模扩大至数吉瓦级，使其 “成为微软各项产品和服务的主要推理系统”。支撑这一计划的架构设计对数据传输机制作出了相当详细的说明，但作为此次发布核心卖点的效率指标，却缺乏同样明确的定义。业内从业者也对此提出了质疑。
目前 Maia 200 已在生产环境中为微软 Copilot 和 MAI 团队提供算力支持。微软计划在今年年底前，将 GitHub Copilot 和 Microsoft Foundry 也接入 Maia 200。该芯片已部署在美国、欧洲和澳大利亚。微软表示，在 Teams 和 Copilot 的内置 AI 功能中，与领先的 GPU 基础设施相比，Maia 200 每瓦可处理的 Token 数量最高可达后者的 1.7 倍。Maia 300 则将于 2027 年开始正式投入生产环境。
Maia 300 的设计以推理为先，主要面向长上下文、内存密集型和解码负载较重的任务，而非训练。值得注意的是，微软的两篇公告都没有提及训练场景。每颗 Maia 300 加速器配备八组 HBM4 堆栈，内存容量为 288 GB，带宽达到 18 TB/s，热设计功耗（TDP）为 950 瓦。一个机架可容纳 96 颗加速器，总功耗为 167 kW，分装在 24 个托盘中，每个托盘包含四颗加速器。整个机架的 HBM 容量达到 27.6 TB，聚合带宽为 1.7 PB/s。
计算 Chiplet 采用台积电（TSMC）的 N2P 工艺制造，内存控制器和 I/O 功能则部署在独立的芯片上。微软表示，这种设计使计算 FLOPs 提升了 36%，同时将内存带宽提高至 Maia 200 的 2.5 倍，并将每瓦 HBM 带宽提升至后者的约两倍。内存访问方式也发生了变化：Maia 300 不再采用统一的内存访问布局，而是让多个 Chiplet 采用 NUMA 架构。这样，软件就可以将数据放置在更靠近实际执行计算的单元的位置。
网络是这套设计中最能体现微软技术取舍的部分。Maia 的 AI 网络控制器实现了一套基于以太网的定制 AI 传输机制。微软将其描述为“一个统一的网络 Fabric，而不是彼此分离的纵向扩展和横向扩展网络域”。加速器可以直接向 Superpod 中的任意对等节点发起通信，“无需经过主机或独立的横向扩展网卡”。
在具体连接方式上，同一托盘内的两颗加速器通过直连链路通信；机架内的交换式网络 Fabric 将全部 96 颗加速器连接起来；第二级光交换机则可将最多 128 个机架互连，组成一个包含 12,288 颗加速器的 Superpod。
微软在文章中划分了三个延迟域：两颗加速器之间的直连通信、Pod 内经过一跳的通信，以及跨 Pod 经过三跳的通信。微软预计，软件会尽可能将张量并行流量限制在延迟最低的通信域内。
机架内带宽与跨机架带宽的分配比例并不会“在系统设计阶段就固定下来”，而是由作业的打包和放置方式决定。因此，当只有部分通信需要跨机架进行时，单颗加速器仍可向任意对等节点提供最高 750 GB/s 的带宽。
这套网络也支撑着 Maia 300 特定的推理执行模式。Prefill 和 Decode 工作节点可以分别确定规模，并独立部署。Prefill 阶段完成后，KV 状态会“通过同一套纵向扩展网络 Fabric 直接传输给选定的 Decode 工作节点，而不是通过另一条由主机中转的横向扩展路径离开当前网络域”。同一套网络还支持跨机架的故障补偿：如果一个机架中的托盘发生故障，另一个机架中的备用托盘就可以接替其工作，整个过程对软件内核透明，无需软件内核感知底层变化。
微软公布的核心数据是，预计 Maia 300“与领先的 GPU 系统相比，每兆瓦处理的 Token 数量将提高 40%”。此外，其机架级 Token 吞吐量最高可达对比系统的七倍，每瓦性能则超过 Maia 200 的 2.5 倍。
真正引发讨论的是这项效率指标。质疑者针对的并非 40% 这个数字本身，而是指标的计量方式和定义。
Rick Bullotta 在 Satya Nadella 的发布公告下发表了互动量最高的评论：
“每瓦 Token 数”简直是个荒谬的指标。仿佛 Token 有统一、一致的定义，或者 Token 数量与计算量之间存在普遍适用的关联。事实并非如此。而且，这种指标本身就会奖励冗长的输出。
评论者 Michel Floyd 则从量纲角度提出了疑问：“指标难道不应该是每瓦每秒处理的 Token 数，或者每千瓦时处理的 Token 数吗？”UBIQS 创始人 Vinay Bansal 正在开发一款与 Maia 竞争的推理加速器。他质疑微软测量这一指标时所采用的工作负载：
每瓦多处理 40% 的 Token，但测试时使用的上下文长度是多少？你们提到的持续运行、多步骤 Agent 任务都涉及很长的上下文，而且每个 Agent 的工作记忆都会随着执行步骤的增加而不断增长。在这种情况下，真正的瓶颈是内存，而不是计算能力。此次发布中最受关注的规格参数，恰恰也是内存规格。
如果这些反馈能够代表业内的普遍看法，那么相比效率指标本身，Maia 300 的架构设计反而更容易评估。微软没有公布作为比较对象的具体 GPU 系统。与此同时，微软自己将 Maia 300 的目标描述为最大限度地提高“每瓦有用 Token 数量”和“每美元有用 Token 数量”，却没有定义什么样的 Token 才算“有用”。两篇公告在表述上也存在一个值得注意的区别：微软执行副总裁 Scott Guthrie 和工程博客都表示，微软预计 Maia 300 的效率将提升 40%；而纳德拉在自己的公告中，则直接将这一提升表述为已经实现的结果。
另一个问题是，谁能使用 Maia。两篇文章提到的所有工作负载都属于微软自身的产品或团队，包括 Copilot、GitHub Copilot、Foundry、MAI 团队和 Dragon Copilot。但两篇文章都没有说明，客户是否能够直接预配 Maia 算力。
Pexon Consulting 首席执行官 Phillip Pham 用德语提出了相关疑问：这些性能提升能否转化为更便宜的预留吞吐量套餐，还是说这类硬件仍将专供微软自己的服务使用？对于采用承诺容量模式的客户而言，真正决定生产环境收益的，是每个有保障 Token 的价格，而不仅仅是硬件本身的处理效率。
除 Rani Borkar 和 Saurabh Dighe 撰写的工程博客外，微软尚未发布其他技术论文。该博客也是此次公告链接的唯一技术资料。
微软表示，将于下个月在旧金山举行的 Microsoft Ignite 2026 大会上公布更多信息。
查看英文原文：Microsoft Commits Maia to Multi-Gigawatt Inference, and Its Efficiency Metric Draws Questions
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Starting in 2027, Microsoft will expand the deployment of the Azure Maia series to the Tgwa level, making it “the main system of reasoning for Microsoft products and services”. The structure that underpinned the plan provided a fairly detailed description of the data transmission mechanism, but the same clear definition was lacking as an indicator of the efficiency of the core selling point of publish. This has also been questioned by practitioners.
Currently, Maia 200 has provided computing support to Microsoft Copilot and MAI teams in the production environment. Microsoft plans to connect GitHub Copilot and Microsoft Foundation to Maia 200 by the end of this year. The chip has been deployed in the United States, Europe and Australia. Microsoft states that, in the built-in AI function of Teams and Copilot, the maximum number of Tokens per watt can be treated by Maia 200 times as high as 1.7 times that of the leading GPU infrastructure. Maia 300 will be officially operational in 2027.
Maia 300 was designed with reasoning in mind and was mainly oriented towards long context, memory-intensive and decoded loads, rather than training. It is worth noting that neither of the Microsoft bulletins mentions the training scene. Each Maia 300 accelerator is equipped with eight sets of HBM4 stacks with a memory capacity of 288 GB, bandwidth of 18 TB/s and heat design power (TDP) of 950 W. A frame can accommodate 96 accelerators, with a total of 167 kW, divided into 24 trays, each containing four accelerators. The entire shelf has a HBM capacity of 27.6 TB and a polymer bandwidth of 1.7 PB/s.
The calculation of Chiplet is made using the TSMC N2P process, with memory controllers and I/O functions deployed on stand-alone chips. Microsoft indicates that this design has increased the calculation of FLOPS by 36 per cent, while increasing the memory bandwidth to 2.5 times that of Maia 200 and increasing the HBM bandwidth per watt to about twice that of the latter. Memory access has also changed: Maia 300 no longer uses a unified memory access layout, but instead allows multiple Chiplets to use NUMA structures. This allows the software to place the data closer to the units actually performing the calculations.
The network is the best part of the design that reflects Microsoft technology trade-offs. Maia 's AI network controller has a customized AI transmission mechanism based on Ethernet. Microsoft describes it as “a unified network, Fabric, not a separate vertical and horizontal extension of the domain”. The accelerator can initiate communication directly to any of the equivalent nodes in Superpod, “without going through the mainframe or an independent horizontal extension card”.
In the specific connection mode, the two accelerators in the same tray communicate through a direct link; the exchange network in the rack, Fabric, connects all 96 accelerators; and the second-stage optical exchange will interconnect up to 128 accelerators into a Superpod containing 12,288 accelerators.
In the article, Microsoft divides three areas of delay: direct communication between two accelerators, one jump in Pod, and three jumps across Pod. Microsoft expects that the software will limit the parallel traffic of the tension to as much as possible within the least delayed communication domain.
The ratio of bandwidth on the rack to the bandwidth across the rack is not " fixed at the design stage of the system ", but is determined by the way the job is packed and placed. Thus, a single accelerator may still provide a bandwidth of up to 750 GB/s to any equivalent node when only part of the communication is required to cross the shelf.
This network also supports Maia 300's specific reasoning execution model. Prefil and Decode work nodes can be scaled and deployed independently. Upon completion of the Prefill phase, the KV status will "transmit directly to the selected Decode working node through the same vertical extension network Fabric, instead of leaving the current network through another horizontal extension path that is being transferred from the host." The same network also supports failure compensation for cross-platforms: if the pallet in one rack fails, the back-up tray in the other rack can take over, the entire process is transparent to the core of the software and no changes in the nuclei of the software are required.
The core data published by Microsoft are expected to increase by 40 per cent the number of Tokens per MW as compared to the leading GPU systems. In addition, the aircraft-grade Token can swallow up to seven times the comparative system and have more than 2.5 times more than Maia 200 per watt.
What really triggered the discussion was this indicator of efficiency. The challenger does not target 40% of this number per se, but rather the way the indicators are measured and defined.
Rick Bullotta published the most interactive comments under the publish announcement of Satya Nadella:
The number of Tokens per watt is an absurd indicator. As if Token had a uniform and consistent definition, or there was a general correlation between the number of Tokens and the calculated amount. That is not the case. Moreover, such indicators themselves reward lengthy outputs.
Commentators, Michel Floyd, ask questions from a quantitative standpoint: “Shall the indicator not be the number of Tokens per second per watt, or the number of Tokens per kilowatt hour? "UBIQS founder Vinay Bansal is developing a competitor accelerator with Maia. He questioned the workload used by Microsoft to measure this indicator:
40% of Token per Vado, but what's the length of context for testing? The ongoing, multistep Agent missions you're referring to involve a long context, and each Agent's working memory will grow as the implementation steps increase. In such cases, the real bottleneck is memory, not computing capacity. The most relevant specifications in this publish are precisely memory specifications.
If these feedbacks represent the prevailing perception in the industry, it is easier to assess the structure of Maia 300 than the efficiency indicators themselves. Microsoft does not publish a specific GPU system as a comparator. At the same time, Microsoft itself described the goals of Maia 300 as maximizing “Token per Wat” and “Token per dollar”, without defining what kind of Token was “useful”. There is also a notable difference in the presentation of the two bulletins: Microsoft’s Executive Vice President Scott Guthrie and the Engineering Blog both stated that Microsoft expected a 40 per cent increase in the efficiency of Maia 300; and Nadra, in his own bulletin, directly expressed this increase as a result of what has been achieved.
Another question is who can use Maia. All of the work loads mentioned in both articles belong to Microsoft's own products or teams, including Copilot, GitHub Copilot, Foundry, the MAI team and Dragon Copilot. However, neither article indicates whether the client can directly plan for Maia's ability to calculate.
Phillip Pham asks the question in German: Can these performance improvements be converted into cheaper pre-input packages, or will such hardware remain dedicated to Microsoft's own services? For customers using a commitment capacity model, what really determines the benefits of the production environment is the price per guaranteed Token, not just the processing efficiency of the hardware itself.
Except for the engineering blog by Rani Borkar and Sarabh Dighe, Microsoft has not yet received other technical papers from publish. The blog is also the only technical resource linked to this announcement.
Microsoft indicated that more information would be made available at the Microsoft Ignite 2026 Conference next month in San Francisco.
See the original English language: Microsoft Committees Maia to Multi-Gigawat Inference, and Its Efficacy Medical Draws Questions
Auto-fetched and translated by AI Atlas. All rights belong to the original author.