It's discussed quite a bit on this site, but I wouldn't expect Apple to pull ahead in the GPU department. They really lit a fire under Intel and AMD's ass with the M1's IPC, but Nvidia laughed their way to the bank with every M1 upgrade. Even the 5nm M1 Ultra struggled to keep up with Nvidia's 10nm, bog-cheap 30-series cards. In the datacenter it's even more of a blowout, Apple would have to invest in something competitive with CUDA to turn heads. That's no small feat, and I don't think it's possible with an overnight API launch. It takes time and integration into the industry, something Apple wasn't patient enough for (see: Xserve).
It's still not favorable, even compared to last-gen (30XX series) hardware. Here's the OpenCL scores for multiple different GPUs, including the Ultra: https://browser.geekbench.com/opencl-benchmarks
The M1 Ultra is a 200w chip, fabricated with 5nm silicon. It's being outclassed by the RTX 3060, a cheap 10nm card that draws ~170w max. In Nvidia's 40-series cards (on 4nm), the M1 Ultra's performance profile is most comparable to a laptop-class RTX 4060 that draws less than 115 watts.
There's a story in the total package draw to be made here, and it does weigh slightly in Apple's favor. Overall though, even with last-gen cards it's clear that there's a massive performance-per-watt lead in Nvidia's favor right now. Which is impressive, considering how they've been stuck with second-class silicon when the M1 Ultra debuted.
https://en.wikipedia.org/wiki/Laconic_phrase
It's discussed quite a bit on this site, but I wouldn't expect Apple to pull ahead in the GPU department. They really lit a fire under Intel and AMD's ass with the M1's IPC, but Nvidia laughed their way to the bank with every M1 upgrade. Even the 5nm M1 Ultra struggled to keep up with Nvidia's 10nm, bog-cheap 30-series cards. In the datacenter it's even more of a blowout, Apple would have to invest in something competitive with CUDA to turn heads. That's no small feat, and I don't think it's possible with an overnight API launch. It takes time and integration into the industry, something Apple wasn't patient enough for (see: Xserve).