Apple was comparing the power envelope (already a complicated concept) of their GPU against a 3090. Apple wanted to show that the peak of their GPU's performance was reached with a fraction of the power of a 3090. What was terrible was that Apple was cropping their chart at the point where the 3090 was pulling ahead in pure compute by throwing more watts at the problem. So their GPU was not as powerful as a 3090, but a quick glance at the chart would completely tell you otherwise.
Ultimately we didn't see one of those charts today, just a mention about the GPU being 50% more efficient than the competition. I think those charts are beloved by Johny Srouji and no one else. They're not getting the message across.
Plenty of people on HN thought that M1 GPU is as powerful as 3090 GPU, so I think the message worked very well for Apple.
They really love those kind of comparisons - e.g. they also compared M1s against really old Intel CPUs to make the numbers look better, knowing that news headlines won't care for details.
They compared against really old intel CPUs because those were the last ones they used in their own computers! Apple likes to compare device to device, not component to component.
Yes, can't remember the precise combo either, there was a solid year or two of latent misunderstandings.
I eventually made a visual showing it was the same as claiming your iPhone was 3x the speed of a Core i9: Sure, if you limit the power draw of your PC to a battery the size of a post it pad.
Similar issues when on-device LLMs happened, thankfully, quieted since then (last egregious thing I saw was stonk-related wishcasting that Apple was obviously turning its Xcode CI service into a full-blown AWS competitor that'd wipe the floor with any cloud service, given the 2x performance)
Gave me quite a laugh when Apple users started to claim they'd be able to play Cyberpunk 2077 maxed out with maxed out raytracing.