Microsoft preparing its own AI chips to compete with Nvidia’s GPUs

Project Athena: AI hardware for OpenAI

The development of artificial intelligence has gained mainstream awareness in recent months with news around ChatGPT and OpenAI and similar projects. These advanced neural networks and AI models have large hardware requirements, benefiting Nvidia, whose GPUs are used to train and run these neural networks. But this interest may also bring new competitors. Among them is reportedly Microsoft, which is preparing its own chips for AI.

This news was reported by The Information website, which cites the testimony of two people allegedly familiar with the Microsoft project. According to them, the software giant is preparing its own hardware adapted to run Large Language Model neural networks, i.e. those capable of generating or “interpreting” texts simulating a human speaker.

Microsoft’s motivation is to get hardware that can handle the extremely demanding large models better than what is readily available on the market (i.e. Nvidia’s GPUs, but also competing specialized accelerators which quite a large number of different companies have already come up with). At the same time, Microsoft also probably wants to achieve savings by not paying Nvidia’s usual margins. Even Microsoft’s various competitors like Facebook or Google (whose TPUs are well known) have similar programs, so this isn’t exactly surprising for Microsoft.

But this is not just a reactionary move due to the current events, although they may have led to accelerating and prioritising the development. Microsoft has reportedly been secretly preparing these AI chips since 2019, and the project, which has the internal designation “Athena”, is at a relatively advanced stage now. According to The Information, the company apparently already has functioning samples, as some Microsoft and OpenAI employees have been given access to them for testing purposes. Perhaps as early as next year, then, the Athena chips could reportedly be made available to a wider range of users within Microsoft and OpenAI, which could mean that some trial production will get underway by then.

This first generation is said to be manufactured on TSMC’s 5nm process, so there are premises for a top-notch level of performance (if the architecture itself is of high quality). Even so, the first generation of Athena chips can probably serve mainly as a test and to gain experience for the next generations that are planned (the company is not working on just one, but on the whole roadmap of gradually improving accelerators). In theory, these could then be made available to external customers, for example as a service in the Azure cloud. But how exactly Microsoft will use its AI hardware commercially is anyone’s guess for now.

Cloud TPU, Google’s AI accelerator offered as a service (source: Google)

The report does not directly mention whether these chips will be used for training neural networks too, or just for inference, i.e. for their application in the Bing search engine and other services. It is possible that in the first phase Microsoft will focus only on inference. The Information mentions that Athena chips are unlikely to completely replace Nvidia’s GPUs at Microsoft, and the company will continue to use those. But Microsoft can probably follow-up with their own accelerators capable of training if the initial scope of the project is a success.

In-house designed chips are not a completely new area of interest for Microsoft. The company was also reportedly interested in developing ARM processors for mobile devices (like the Surface products) and servers. However, no tangible results that would be comparable to Amazon’s well-known Graviton processors have yet publicly emerged on that front.

Sources: The Information, The Verge, Tom’s Hardware

English translation and edit by Jozef Dudáš

Flattr this!

Leave a Reply

Your email address will not be published. Required fields are marked *