微软新一代 AI 芯片降低对 Nvidia GPU 的依赖

微软今日宣布推出第三代自研 AI 芯片 Maia，该芯片旨在成为微软产品和服务的“主要 AI 推理系统”。Maia 300 不仅能够改善云端运行 AI 工作负载的成本与效率，也更能胜任最新的推理模型和 Agent 任务。

Maia 300 将于 2027 年开始投入生产部署，其设计重点是效率。“我们预计，与领先的基于 GPU 的系统相比，Maia 300 每兆瓦可提供多 40% 的 token，”微软云业务集团执行副总裁 Scott Guthrie 表示。“在完整的 96 加速器配置中，它的机架级 token 吞吐量最高可达 Maia 200 的 7 倍。”

这些效率提升将使微软能够在不久后主要依赖自研 AI 芯片，而不是 Nvidia 和 AMD 的芯片。“我们将从 2027 年起把 Azure Maia 产品家族扩展至数吉瓦级部署，使 Maia 成为微软产品和服务的主要 AI 推理系统，”Guthrie 表示。

Maia 300 是 Maia 200 的后继产品，后者仅在今年 1 月宣布。Maia 200 已在微软的生产级 AI 工作负载中运行，包括 Microsoft Copilot、GitHub Copilot、Microsoft Foundry，以及微软自有的 AI 模型。

“以 AI 服务的真实世界指标衡量，Maia 200 是我们设备集群中性能最佳的 AI 基础设施，”Guthrie 表示。“Maia 300 是一套完整的 AI 系统，专为当下正在涌现的新型 AI 工作负载设计：长上下文推理、对延迟敏感的对话体验，以及持续执行、多步骤的 Agent 任务。”

微软正采用新的 chiplet 架构来实现至关重要的效率提升。最新的 Maia 300 架构基于台积电 N2P 制程节点打造，将计算、内存控制器和 I/O 功能分离。“这增加了专门用于计算性能的硅片面积，并使每项功能都能够采用最适合其职责的技术，”Azure 硬件与基础设施团队表示。

Maia 300 也是微软自研 AI 模型工作的重要组成部分，并可能帮助该公司实现追赶 Google DeepMind、OpenAI 和 Anthropic 的目标。微软 AI CEO Mustafa Suleyman 表示：“我们必须覆盖整个技术栈，一直深入到硅片层面。在这一规模下，我们需要一个由自己拥有、并与我们的模型协同设计的平台。这正是让微软 AI 模型运行在 Maia 芯片之上，成为我们团队关键重点的原因。”

Microsoft is announcing its third-generation Maia in-house AI chip today, designed to be the “primary AI inference system” for Microsoft’s products and services. Maia 300 not only improves the cost and efficiency of running AI workloads in the cloud, it’s also more capable of handling the latest reasoning models and agentic tasks.

Maia 300 will start being rolled out to production in 2027, and it has been designed with efficiency in mind. “We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems,” says Scott Guthrie, executive vice president of Microsoft’s cloud group. “And in a full 96-accelerator configuration, it delivers up to 7x the rack-scale token throughput of Maia 200.”

These efficiency improvements will allow Microsoft to rely mainly on its in-house AI chips soon, instead of ones from Nvidia and AMD. “We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary AI inference system for Microsoft products and services,” says Guthrie.

Maia 300 is a successor to the 200, which was only announced in January. Maia 200 is already running on production AI workloads across Microsoft, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and Microsoft’s own AI models.

“Maia 200 is the best performing AI infrastructure in our fleet, measured in real world metrics for AI services,” says Guthrie. “Maia 300 is a complete AI system designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks.”

Microsoft is using a new chiplet architecture to get its all-important efficiency gains. Built on TSMC’s N2P process node, the latest Maia 300 architecture separates compute, memory controller, and I/O functions. “This increases the silicon area dedicated to compute performance and allows each function to use the technology best suited to its role,” says the Azure hardware and infrastructure team.

Maia 300 is also a key part of Microsoft’s own AI model work, and could help the company achieve its goal of catching up with Google DeepMind, OpenAI, and Anthropic. “We have to work across the full stack, all the way down to the silicon,” says Microsoft AI CEO Mustafa Suleyman. “At that scale we need a platform that we own, and that we co-design with our models. That’s why running our Microsoft AI models on top of the Maia chip is a critical focus for our team.”