Apple may be late to the generative AI party, but don’t count it out just yet. According to Bloomberg‘s Marc Gurman and MacRumors‘ Hartley Charlton, the company will be the M2 Ultra in its own servers – in its own data centers – to support its growing GAI ambitions. Launched in June 2023, the CPU – as used in the Mac studio – remains the most complex piece of silicon ever released by Apple, with 24 computing cores, up to 76 GPU cores and 32 AI accelerators.
The report does not say whether Apple plans to revive its defunct Xserve line of rack servers, nor whether its Mac OS X server operating system. Both products have been on hold for years, because Apple shifted its focus from the business market at the beginning of the last decade. A separate article from WJ also adds that Apple uses the internal codename ACDC (Apple Chips in the Data Center).
In this piece, authors Aaron Tilley and Yang Jie argue that Apple would use the formidable firepower of its data centers for training or more complex inference, while lighter workloads (or those that require access to personal data) would be handled locally on the device. itself, eliminating the need to run in the cloud.
This echoes what chip makers like x86 giants AMD and Intel, in collaboration with Microsoft, have been advocating with the AI PC paradigm: large server chips (such as Epyc And Xeon) work together with smaller client processors (Ryzen or Core). The difference, of course, is that Apple uses an existing processor instead of a new one.
Another way to achieve AI hegemony?
That raises another question; Did Apple plan this from the beginning, with a do-it-all CPU family? Bearing in mind that the M2 Ultra would probably be the only server processor in the world with a GPU and an AI engine. Could it make way for a server-only version (the S1?) aimed at a data center environment, with many more cores, no GPU and much, much more memory?
All in all, however, there was never any doubt that Apple would sooner or later get involved in server processors. It was a matter of when, not if. Reports that Apple is building its own servers date back to 2016 and is in line with Apple’s doctrine of owner of the stack. In 2022the company also wanted to recruit a “fun-loving and hard-working hardware validation engineer” to “develop, implement and complete hardware validation plans for the next generation of hyperscale and storage server platforms.”
Then, a year later, research conducted by analyst firm Structure Research found that Apple planned to triple the critical power capacity of its data centers to accommodate its two billion active devices (and nearly a billion iOS users) and deliver more services.
Of course, hardware requires software, and Apple has been increasingly vocal in the past twelve months by releasing MLX, a machine learning framework designed specifically for Apple Silicon. AI-enhanced Siri and a new set of AI tools called OpenELMs.
It will be extremely educational to see how Apple manages to do generative AI at scale using something other than brute force GPU (à H100). This could have a direct impact on the fate of another trillion-dollar company called Nvidia). WWDC, the annual Apple developers conference, takes place next month and it will have AI written all over it.