Abstract:
After years of relying on Nvidia graphics processors, Microsoft will expand the use of self-developed artificial intelligence chips as the software giant aims to strengthen its full-stack AI capabilities. Microsoft officially released the third-generation Maia AI acceleration chip on Monday and said that subsequent processing of business requests will mainly rely on its own chip hardware.
Microsoft's biggest advantage in the field of AI is the Azure cloud platform, which runs various AI models for customers. Microsoft 365 Copilot and GitHub Copilot provide AI services for programmers and office workers. Microsoft is also developing large self-developed models, but in a market dominated by OpenAI, Anthropic, and Google, self-developed models have yet to have significant influence.
The chip track is ushering in new opportunities, stemming from the market's continued strong demand for high-end processors and the market's demand for diversified chip choices. NVIDIA GPUs have always been the industry standard for AI computing workloads and model training. Microsoft's existing data centers use this product extensively, and are also used to run OpenAI and Anthropic models.
However, as demand continues to rise, leading model companies have begun to use chips from multiple manufacturers, including Amazon, AMD, Cerebras, Google, and in recent months, Microsoft's self-developed chips.
In the general computing scenario, Microsoft already uses its own CPUs as well as chips from Intel and AMD. Chief Financial Officer Amy Hood said at a Goldman Sachs conference last month that self-developed chips have helped the company
profit margin improvement
over the past decade. .Microsoft’s first AI chip, Maia 100, will be launched in 2023; Maia 200 will be released later. CEO Satya Nadella told analysts on an earnings call that the Maia 200 "provides 30% more achievable performance per dollar than existing hardware," and that the chip is already used to host OpenAI models.
Series production of the Maia 300 will begin in 2027. Microsoft Executive Vice President Scott Guthrie said in a blog that the goal is to achieve a multi-gigawatt deployment scale for the Maia series of chips, but General Manager Alistair Spears did not give a clear implementation timetable.
Spears said: "We retain a certain amount of room for choice. This industry iterates very quickly, and we hope to match the most suitable chip for the corresponding model at the right time."
Microsoft plans to deploy 96 Maia 300 chips per data center cabinet, twice the number of Maia 200. The AI token processing throughput of a single cabinet can be up to 7 times that of the previous generation; compared to GPUs, the number of tokens produced per megawatt of energy consumption is increased by 40%.
“We need a new architecture to achieve better performance and improve power efficiency. For data centers, power consumption directly corresponds to operating costs.” Spears said.
Microsoft Technical Academician Jim Clevein said that some Microsoft 365 Copilot services can be accelerated with the help of Maia 300.
In a research report sent to clients on Wednesday, analysts from investment bank Evercore ISI estimated that Maia chips will account for 10% of Microsoft's AI chip procurement expenditures by the first quarter of 2027; the institution gave Microsoft stock a buy rating.
Nvidia will remain an important partner of Microsoft. Last week, Microsoft held a product launch event, and Nvidia CEO Jensen Huang was present. At the meeting, Microsoft released a new Surface notebook equipped with Nvidia processors. Spears also mentioned that Microsoft will still cooperate with Nvidia and AMD in the model training business.
Unlike Google TPU, Amazon Inferentia and Trainium chips,
Microsoft's acceleration chip is not currently available for rent to other companies on an hourly basis
.Patrick Moorhead, CEO of Moor Insights & Strategy, an industry consulting firm, said: "At present, no external company has tested this chip except within Microsoft."
Related articles:
Microsoft releases the third-generation self-developed AI chip Maia 300, which increases the computing power efficiency per megawatt by 40%
Comments