Microsoft desires to launch an AI chip because it appears to place itself because the Apple of AI, providing a vertically built-in resolution for scale-out synthetic intelligence. In response to the Info (opens in new tab), Microsoft has been engaged on the chip, codenamed Athena, since 2019 and engineering samples are already accessible to some Microsoft and OpenAI staff. 2019 was the yr when the previous invested $1 billion within the latter; so there’s a possible hyperlink between the plan for the chip and this primary funding.
Different US primarily based corporations (Amazon, Google, Fb) and non-US entities (Tencent, Alibaba) have launched their very own custom-built chips for AI with the intention to get a {hardware} resolution that performs higher than what’s commercially accessible off the shelf. These processors are important for inference and large-language mannequin (LLM)coaching and proper now, Nvidia corners the market with the A100 and the H100.
In doing so, they’re mimicking Apple’s uber-successful vertically built-in, hardware-and-software stack, that’s deployed throughout billions of units worldwide. Microsoft unsuccessfully tried to observe Apple’s footsteps within the client market with the ill-fated Home windows RT on Arm virtually a decade in the past however this time round, Microsoft appears to be extra in management, and most significantly, not a laggard.
Saving billions
Microsoft has, to place it mildly, embraced synthetic intelligence at scale with enthusiasm. It has baked ChatGPT throughout all its flagship services and products (Bing, Microsoft 365, Github) and it’s solely a matter of time earlier than it will get deployed throughout its whole Azure stack. The largest impediment to the rollout is the {hardware} price which, in response to consultants, might attain tens of billions of {dollars} per yr. Microsoft is already collaborating with Nvidia to construct a next-generation supercomputer and it’s probably that the world’s largest software program firm might leverage Athena to get higher charges or precedence from Nvidia sooner or later.
Apart from the truth that about 300 individuals are engaged on Athena at Microsoft, the Info additionally claims that the roadmap for the chip contains “a number of generations” and the primary technology might be fabbed utilizing TSMC’s 5nm manufacturing course of. Ex-Apple CPU architect, Mike Filippo, is prone to have a hand in that venture though his experience lies extra with CPU.
To some extent, Challenge Athena is a observe as much as Challenge Brainwave, the FPGA-powered system that Microsoft has been utilizing since 2017 when it began dabbling with AI. Again then, Intel Stratix 10 FPGA (Subject Programmable Gate Array) playing cards had been used to supply what now appears like the primary draft of Microsoft AI technique. Ought to Microsoft observe Google’s step with a TPU-like product, we are able to anticipate a variety of high-bandwidth reminiscence (HBM), a comparatively low TDP and a laser-focus on peak FLOPs per greenback.