Microsoft will unveil its first proprietary processors during Ignite 2023: Azure Maia and Azure Cobalt. The processors are designed to support AI and Azure services in data centers.

Microsoft sees Azure Maia and Azure Cobalt as the final piece of the puzzle. The processors are expected to roll off the assembly line from the beginning of 2024. Microsoft will house the processors in its own data centers. Production will be outsourced to TSMC, but Microsoft will make custom server racks for the chips itself.

Maia and cobalt

Azure Maia is an “AI accelerator” designed for Windows Copilot and Bing Chat. Since the design of the chip is also tailored to the hardware in the data centers, this should lead to serious performance and efficiency gains, writes Microsoft in a blog. Microsoft used ARM architecture for Azure Cobalt. This chip is intended to make Azure cloud services more efficient overall and increase the “performance per watt” of each server.

In addition, Microsoft remains rather reserved when it comes to the specifications of the chips. The press release makes no mention of cores, flops, memory or bandwidth and no benchmarks are available yet. We have to be content with a statement from OpenAI CEO Sam Altman, who is of course very enthusiastic about the processors. A comparison of Maia with Nvidia’s newly announced H200 Hopper is therefore not possible.

Microsoft will primarily use Maia and Cobalt internally, although Azure customers can also use them to deploy their heavy workloads. Microsoft is also adding another one Azure Boost a system that accelerates storage and networking.

What you do yourself, you do better

The self-developed processors are of great strategic importance for Microsoft. The software giant has been developing its own hardware for its data centers since 2016. Processors fit perfectly into these plans. Now Microsoft has processors optimized for Azure cloud services.

Microsoft also wants to become more independent from external parties. AI services like Windows Copilot and Bing Chat cost Microsoft tons of money because it has to buy expensive processors from Nvidia. By developing the AI ​​chips yourself, you can save a lot of costs.

That doesn’t mean Microsoft will suddenly cut all ties with Nvidia. Microsoft announces virtual machines tailored to the H100 Hopper GPU and will offer the new generation H200 via Azure starting next year.