Microsoft 宣布 Maia 300 AI芯片，多吉加瓦 Azure 缩放

AI 模型和平台

Microsoft 宣布 Maia 300 AI芯片，多吉加瓦 Azure 缩放

在 Google 上的首选来源中添加 Unite.AI

Microsoft于2026年10月12日推出 Azure Maia 300，是其内部AI加速器家族的第三代，称将自2027年起将Azure Maia家族规模化为多基瓦部署，使Maia成为其产品和服务的主要推论系统。

Scott Guthrie),Cloud + AI的执行副总裁，宣布的芯片在官方的Microsoft博客文章中描述Maia 300是一个"完全的智能系统"，旨在进行长文推理，耐久敏感的对话体验和持续，多步骤的代理任务。 Maia 300将在2027年开始生产部署。

Microsoft通过提高整个系统的效率，预计Maia 300将比主要基于GPU的系统每兆瓦多交付40%的代币，Guthrie写道。 在完整的96加速器配置中，称公司，Maia 300可提供高达7x的马亚200的架式符号吞吐量。

Maia 300系统架构

Maia 300的设计采用了Microsoft所称的推论第一架构优化，用于长文，内存密集和解码重的AI工作量，一种公司称涉及在云尺度上平衡计算，内存带宽，联网和功率效率的方法。 其软件定义的数据流架构与AI辅助内核优化对齐，以保持数据本地化，剪切数据移动，提高跨计算，内存和网络资源的利用率。 Microsoft 称将数据流作为一流设计原则处理，使Maia 300比传统的GPU架构具有显著的每瓦性能优势。

内存系统使用8个HBM4堆栈，在950瓦的TDP内发送18.0 TB/s的峰值带宽，专门的内核支持将die SRAM作为软件管理的刮板处理，以减轻对HBM的压力。 芯片也转向多芯片设计，将计算器，内存控制器和I/O函数分开，由TSMC的N2P工艺节点所建的计算芯片组成。 支持高级低精度格式，包括增强的MXFP4，旨在减少计算和KV-cache存储的内存和带宽要求，使上下文windows和同一功率信封内更高的吞吐量成为可能。

在网络方面，Maia 300将网络视为其数据流架构的延伸，而不是通过单独的通信层移动数据。 加速器之间的高带宽互联被优化为拉伸平行性，定制AI运输在以太网为基础的基础设施上提供加速器和架之间的规模化通信，形成从一个单架到数据中心规模部署的统一的地形。 在机架一级，96 Maia 300加速器在一个167千瓦的电源包内运行，在扩容域内有27.6TB的HBM容量和1.7PB/s的总HBM带宽。 密度的Microsoft 称使得它能够服务于更同步的AI交互，以及每个机架更大的工作量。

Maia 200从事生产

Maia 300跟随Maia 200，它已经服务于Microsoft的生产工作量，包括Microsoft Copilot，并支持MAI团队开发下一代型号，GitHub Copilot和Microsoft铸币局将在2026年年底前跟进。 在这些工作量中，芯片运行着OpenAI模型，Microsoft自己的MAI模型，并带领开源模型家庭。 Maia 200正在美国、欧洲和澳大利亚各地部署，Microsoft 称，它为Teams、Copilot和Cowork经验的内置AI能力每瓦多提供1.7x的代币，与主要的GPU基础设施相比。

Guthrie写道，在Maia 200上运行Microsoft的AI模型和服务的经验直接告知了Maia 300的设计。 他的称，两代设计都是为了支持一套广泛的额外的Microsoft AI服务：GitHub编码经验背后的多步代理工作，云尺度模型通过Microsoft创建服务，推论和为MAI团队生成数据，以及旨在减轻临床医生行政负担的耐久感龙Copilot工作流程。

Microsoft 推出 Maia 200于2026年1月26日，作为TSMC的3纳米计工艺上编造的推论加速器，具有原生的FP8和FP4拉伸芯，216GB的HBM3e内存为7TB/s,272MB的芯片SRAM. 芯片包装了超过1400亿个晶体管，在一个750瓦的TDP信封内，在FP8上交付了超过10个PetaFLOPS的FP4计算和超过5个PetaFLOPS. 其两级扩增网络以标准以太网为基础，提供每个加速器的2.8 TB/s双向扩增带宽，跨越高达6,144个加速器的集群。 Maia 200最早部署在爱荷华州德斯莫因斯附近的Microsoft的美国中央数据中心区域，美国西部3区在菲尼克斯附近，亚利桑那州宣布作为下一个区域。

在发布时，Microsoft将Maia 200描述为来自任何超标器的最具表现力的第一党硅，引用了第三代Amazon列车的FP4性能的3倍，超过Google第七代TPU的FP8性能，每美元性能比机队最新硬件好30%. 该公司还称马亚200是它所部署的最有效的推断系统。

Microsoft 称将在旧金山的Microsoft Ignite 2026和网上分享其战略和前沿转型进展，并将读者指向Maia 300技术论文，以进一步了解建筑细节。

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。

AI Models & Platforms

Microsoft Announces Maia 300 AI Chip, Multi-Gigawatt Azure Scale-Up

Add Unite.AI to your preferred sources on Google

Microsoft on October 12, 2026 introduced Azure Maia 300, the third generation of its in-house AI accelerator family, and said it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system across its products and services.

Scott Guthrie, executive vice president of Cloud + AI, announced the chip in an Official Microsoft Blog post, describing Maia 300 as a “complete intelligent system” designed for long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks. Maia 300 will begin production deployments in 2027.

By improving efficiency across the system, Microsoft expects Maia 300 to deliver 40% more tokens per megawatt compared with leading GPU-based systems, Guthrie wrote. In a full 96-accelerator configuration, the company said, Maia 300 delivers up to 7x the rack-scale token throughput of Maia 200.

Maia 300 System Architecture

Maia 300 is designed with what Microsoft calls an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads, an approach the company says involves balancing compute, memory bandwidth, networking and power efficiency at cloud scale. Its software-defined dataflow architecture pairs with AI-assisted kernel optimization to keep data local, cut data movement and raise utilization across compute, memory and networking resources. Microsoft says treating dataflow as a first-class design principle gives Maia 300 a significant performance-per-watt advantage over traditional GPU architectures.

The memory system uses eight HBM4 stacks delivering 18.0 TB/s of peak bandwidth inside a 950-watt TDP, and specialized kernel support treats on-die SRAM as a software-managed scratchpad to ease pressure on HBM. The chip also moves to a multi-chiplet design that splits compute, memory-controller and I/O functions apart, with the compute chiplets built on TSMC’s N2P process node. Support for advanced low-precision formats, including enhanced MXFP4, is designed to reduce memory and bandwidth requirements for both computation and KV-cache storage, enabling larger context windows and higher throughput within the same power envelope.

On the networking side, Maia 300 treats the network as an extension of its dataflow architecture rather than moving data through separate communication layers. High-bandwidth interconnects between accelerators are optimized for tensor parallelism, and a custom AI transport provides scale-up communication across accelerators and racks over Ethernet-based infrastructure, forming a unified topology that scales from a single rack to datacenter-scale deployments. At rack level, 96 Maia 300 accelerators operate within a 167-kilowatt power envelope, with 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth across the scale-up domain. Microsoft says that density allows it to serve more simultaneous AI interactions and larger workloads from each rack.

Maia 200 in Production

Maia 300 follows Maia 200, which is already serving Microsoft production workloads including Microsoft Copilot and supporting the MAI team as it develops next-generation models, with GitHub Copilot and Microsoft Foundry to follow by the end of 2026. Across these workloads, the chip runs OpenAI models, Microsoft’s own MAI models and leading open-source model families. Maia 200 is deploying across the United States, Europe and Australia, and Microsoft says it is delivering up to 1.7x more tokens per watt for built-in AI capabilities in Teams, Copilot and Cowork experiences compared with leading GPU infrastructure.

Guthrie wrote that experience running Microsoft’s AI models and services on Maia 200 directly informed the design of Maia 300. Together, he said, the two generations are designed to support a broad set of additional Microsoft AI services: multi-step agentic work behind GitHub coding experiences, cloud-scale model serving through Microsoft Foundry, inference and data generation for the MAI team, and latency-sensitive Dragon Copilot workflows intended to reduce clinicians’ administrative burden.

Microsoft introduced Maia 200 on January 26, 2026, as an inference accelerator fabricated on TSMC’s 3-nanometer process with native FP8 and FP4 tensor cores, 216GB of HBM3e memory at 7 TB/s and 272MB of on-chip SRAM. The chip packs more than 140 billion transistors and delivers over 10 petaFLOPS of FP4 compute and over 5 petaFLOPS at FP8 within a 750-watt TDP envelope. Its two-tier scale-up network is built on standard Ethernet, providing 2.8 TB/s of bidirectional scale-up bandwidth per accelerator across clusters of up to 6,144 accelerators. Maia 200 was first deployed in Microsoft’s US Central datacenter region near Des Moines, Iowa, with the US West 3 region near Phoenix, Arizona announced as the next region.

At that launch, Microsoft described Maia 200 as the most performant first-party silicon from any hyperscaler, citing three times the FP4 performance of the third-generation Amazon Trainium, FP8 performance above Google’s seventh-generation TPU, and 30% better performance per dollar than the newest hardware in its fleet. The company also called Maia 200 the most efficient inference system it had ever deployed.

Microsoft said it will share more on its strategy and Frontier Transformation progress at Microsoft Ignite 2026 in San Francisco and online, and pointed readers to a Maia 300 technical paper for further architectural detail.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.