Microsoft is secretly working on its own artificial intelligence (AI) chips, named "Athena," to mitigate the high costs of development for both its in-house and OpenAI projects. The move is seen as an attempt to adopt Apple's successful hardware-and-software stack approach. The project has been underway since 2019, and engineering samples are already available to select Microsoft and OpenAI employees.
The high cost of GPUs, such as Nvidia's H100, has pushed several tech giants, including Meta, Google, and Amazon, to develop their own machine learning chips. By creating custom-built chips for AI, these companies aim to achieve better performance than what is available off the shelf. Microsoft's AI chip project is anticipated to save billions of dollars per year in hardware costs.
Currently, Nvidia dominates the market with its A100 and H100 GPUs. Microsoft's partnership with Nvidia on a next-generation supercomputer could potentially leverage Athena to negotiate better rates or priority from Nvidia in the future. The Information reports that about 300 people are working on Athena at Microsoft, and the first-generation chips will be manufactured using TSMC's 5nm process.
Project Athena follows Microsoft's Project Brainwave, which began in 2017 and utilized Intel Stratix 10 FPGA (Field Programmable Gate Array) cards for AI applications. If Microsoft develops a TPU-like product similar to Google's, it will likely feature high-bandwidth memory, a relatively low TDP, and a focus on peak FLOPs per dollar.