The new AI arms race at technology giants like Google and Microsoft looks to be forcing companies to go to new extremes. In the latest, according to a report from The Information (paywalled), as spotted by The Verge Microsoft is already working on its own AI chips that can be used to train its products, in an effort to reduce dependence on Nvidia's own server chips.
Apparently codenamed as "Athena" this Microsoft AI chip has been in work since as early as 2019 and is already in use by select Microsoft and OpenAI employees, with a broader internal rollout possibly coming next year. It's not clear, however, if the chips will remain as something that will be for internal use only or will also ship out to its own customers who use Azure cloud products. There are also several future versions of the chip planned, though The Information and The Verge note that the chips aren't something that will directly replace Nvidia's own H100 GPUs.
A report from a few years ago in 2020 indicated that Microsoft was considering its own custom ARM-based chips for servers, but it's not clear if this is related to Athena at all. Microsoft already has its own SQ series of Arm-based chips in Surface devices, which it worked together with Qualcomm on. It also worked with AMD on custom Ryzen CPUs for the Surface Laptop series of devices. So, Microsoft clearly has experience in this chipset area as it continues to push Artificial Intelligence across its core products.
This move, shouldn't be too shocking for those who follow the industry. Google uses its own AI Supercomputers which have over 4.000 TPUs and custom parts that can run AI models. The full documentation was recently revealed online and this was in use since 2020. Amazon also has its own Trainium chips which are used to train machine-learning models in natural language processing. Meta, meanwhile, uses Nvidia and AI chips in its supercomputer, the Research Super Computer.