Microsoft has spent billions of dollars on AI software, like building out its Azure virtual machine infrastructure and putting ChatGPT into products like Bing and SwiftKey. In 2019, there were rumors that Microsoft was getting into AI hardware by making a chip that was designed to work with AI. Now, The Information says engineering samples of the new chip, which has the codename Athena, have been sent to some Microsoft and OpenAI workers.
The story (which you have to pay to read) says that Athena will use TSMC’s 5-nanometer process, at least for its first version. At the moment, most, if not all, high-performance silicon is made at the 5nm scale.
Both AMD and Apple use TSMC’s 5nm technology. But The Information also says that the Athena roadmap has multiple versions,’ which could include TSMC’s upcoming 3nm process.
Microsoft could save money by making its own chips, like Google’s tensor processing units (TPUs) or Amazon‘s Graviton 3. It is said that Athena will cost Microsoft about $100M. At the moment, Nvidia has most of the GPU market, and Microsoft’s Azure data centers already use a lot of Nvidia hardware. Dylan Patel of the research company SemiAnalysis told The Information, “If Athena is competitive, it could cut Nvidia’s chip prices by a third.”
The fact that Microsoft is rumored to be bringing its AI chip creation in-house is more proof of a long-term shift in how chips are made. For decades, hardware companies like AMD, Intel, and Nvidia sold their designs to businesses like Microsoft and Google.
Most companies still do this, but the biggest ones, like Apple, Microsoft, and Google, now create at least some of their own hardware instead of relying on GPU or CPU companies to move the market forward.
Vertically linked hardware and software design could also help Microsoft make its code run faster. Even though the companies work together, the designs of Nvidia’s silicon and Microsoft’s planned data flow aren’t exactly the same.
And it costs $700,000 every day to keep ChatGPT working. Patel said that Microsoft spends about $0.36 on each ChatGPT question. With purpose-built silicon, Microsoft could save about $84 million, which would cover almost all of the cost of getting Athena up and running.
Microsoft may be trying to avoid the “Nvidia tax” by making its own chips for AI processing, but Athena won’t be the last time the two companies work together. Microsoft will still be able to use Nvidia tools for AI projects.
The two companies are working together on a next-generation AI system that would use ‘tens of thousands’ of Nvidia’s A100 and H100 GPUs and Microsoft’s Azure virtual machines.