Sure. But you don't need to go to 3-4 GPU's to see a big gap. Let's consider just dual-GPU machines. When the AS MacPro comes out, it's going to be competing with dual 4090 (est. 160 TFOPS), dual 4090 Ti (est. dual 180 TFLOPS) or dual Quadro cards with whatever the sucessor is to the A6000 (comparable to the 4090/4090Ti).
Indeed, an M2 Extreme (est. 60 TFLOPS, unless they do something fancy) isn't even going to be competitive with a single 4090 (est. 80 TFLOPS) in general GPU compute.
Now maybe it doesn't need to be. But let's be accurate in characterizing where the gap will actually be. It's won't be vs. 4 X GPU machines. That's a straw man.
So the argument you'd actually need to make isn't that only a tiny fraction of the workstation market is running 4x GPUs. Rather, you'd need to argue that only a tiny fraction of the workstation market will run anything more powerful than a single 4080 Ti (est. 60 TFLOPS) (or the Quadro equivalent).
So I'm setting up a straw man ( when I did not come up with the 4x GPU metric in the first place) and you are not even following your single GPU comparison ( 'competing with dual 4090 ...' ) advice at the top of your response. Yes, if look at where the bulk of the workstation market is now, the upcoming upper-high end range cards are mostly gong to replace single and mid-range duals. The highest upper end multiple GPU set ups are going to drift into another 'zipcode' of the more fringe workstation market. Apple doesn't need to go there to sell enough units to have a viable product.
Nvidia is chopping the power to the "Pro" / Quatra class variant of the 4090's die down 150W.
" ... has a TDP rating of up to 300W power (vs. 450W in case of the GeForce RTX 4090). ..."
Nvidia unveils new ProViz flagship: Old name, new board.
www.tomshardware.com
Somewhat doubtful the RTX 6000 will be pulling SP 80 TFLOPs when the power budget is cut. It will not be all the way back to the Extreme's theoretical upper limit, but closer.
A large chunk of the "Pro" value add is boosting the GPU VRAM capacity up to 48GB of RAM. The 'minimal' memory for a 'Extreme' is likely going to be double that of a Ultra (min 64GB ) . So starting at 128GB. Even if split that 60:40 CPU:GPU that is 3GB more VRAM for the "Extreme". If split 50:50 an even bigger gap.
Dual 3060-3070 deployments likely greatly outnumber of the number the old dual 3080/3090 deployments. The 4090 is self restricting card. The power and volume requirements bump it out of many systems (even out of the MP 2019 chassis).
Have front
and back side water blocks.
EK GOES ABOVE AND BEYOND TO ROLL OUT WATER BLOCK FOR JUST RELEASED NVIDIA RTX 4090 GPUS The launch of NVIDIA® GeForce® RTX™ graphics cards’ latest edition for gamers and content creators is finally announced. The GeForce RTX 40 Series will be the 3rd generation of RTX graphics cards. Expected...
videocardz.com
> 3 width cards.
"... Strix GeForce RTX 4090, we built it with cooling as our top priority. The card’s 3.5-slot design consists of a ...
...
... the new ROG Strix GeForce RTX 4080 16GB ...In order to do that, this 3.5-slot cooler design uses the same die-cast metal frame, shroud, and backplate as the ROG Strix GeForce RTX 4090...
...
UF Gaming GeForce RTX 4080 16GB brings next-gen performance in a sturdy design. ... with a 3.65-slot thick heatsink for next-level cooling. That’s slightly thicker than the Strix variants of this card, but also measures more than 9mm shorter for wider case compatibility. ..."
ASCEND TO NEW HEIGHTS WITH THE NEW GEFORCE RTX 4090 AND RTX 4080 CARDS FROM ROG STRIX AND TUF GAMING NVIDIA’s long-awaited GeForce RTX 4000 series of graphics chips is finally here, and PC builders will soon be able to get their hands on a new series of gaming cards—starting with the top-tier...
videocardz.com
Nvidia recommendation for the 4090 is a 700W power supply. Dual 4090s would consume 100% of the whole MP 2019 power supply just by themselves. (***) It will blow out the power budget for a large fraction of Dell/HP/Lenovo workstations also. ( that is one reason why the Pro model doesn't try to go anywhere near that high. )
The 4090 is a funhouse mirror card, whose primary objective is to claim the kind of tech porn benchmark crown. Nvidia fanboys are going to throw many at it. Folks past eyeball deep in highly proprietary Nvidia libraries will throw money . But it isn't a practical card that is going to sell in relatively high numbers.
The 4080 configurations do some very substantive backoffs to get to more pragmatically more sane system constraint parameters.
4090 24GB 450W 128SM 356-bit bus 1008GB/s memory bandwidth
4080 16GB 320W 76SM 256-bit bus 736GB/s memory bandwidth
4080 12GB 285W 60SM 192-bit bus 504GB/s memory bandwidth
NVIDIA debuts “Ada” GeForce RTX 40 series NVIDIA formally introduces three new graphics cards. The new RTX 40 series are here. NVIDIA CEO Jensen Huang has now officially introduced RTX 4090, RTX 4080 16GB and RTX 4080 12GB models, the first SKUs from the next-generation of graphics series...
videocardz.com
By time dropped back to power levels more comparable with what Apple is going to consume the memory bandwidth has dropped in half. Those theoretical peak SP TFLOPs are less likely to match on more than few apps. Nvidia has some bigger caches which will help with micro benchmarks, but for moving data back and forth; not quite as much.
The bigger SP TFLOPs on broad spectrum apps 'threat' is likely going to come from AMD not Nvidia. AMD likely will hold back at first but at some point seems likely to just top the the 4090 in those cases with a 7900 with extra Infinity cache. (same central compute die , but 'bigger' memory controller + cache chiplets ) No good reason to ship those at first ( until they can measure exactly what the 4090 does) , but it is apparent that Nvidia has clocked the 4090 way into the diminishing returns zone. The weaker underbelly of Nvidia is going to be the higher end laptop dGPU space which is where there is a much better opportunity over next 18 months to make really good progress on unit sales.
The only thing that might 'scare' Apple about the 4090 is the 3rd generation hardware RT stuff. The out of control thermals. The half memory bandwidth at reasonable card parameters . That stuff isn't going to send Apple quaking in a corner in fear for the upcoming Mac Pro. Folks eyeball deep in Nvidia proprietary software... there are no Nvidia drivers even on macOS Intel. Those folks aren't causing Apple to loose tons of sleep at night. Could be nice to have those folks, but not critical to the Mac ecosystem. Apple did the Plus/Minus trade-off evaluation a while back during their feud with Nvidia over GPU driver priorities and requirements.
What would be useful for Apple to have is something like
" ... Peak Single Precision (FP32) Performance
22.6 TFLOPs ..."
https://www.amd.com/en/products/server-accelerators/amd-instinct-mi210"...
Which could give some configurations another booth in TFLOPs without some crazy high TDP increase. It is just a 'compute' card that could fit in some expansion box across the Mac line up. Short term, I doubt it is coming because they really haven't laid the ground work for it. But slavishly tracking where there 4090 is mainly going... extremely doubtful Apple is going to follow that route.
(***) Yes part of Nvidia's 700W budget is covering some modest allocation for "&rest" of the PC that is needed to run to 4090. But is indicative that the card is trying to 'hog' as much of the budget as it can. Peak, spike loads are likely going to be relatively high over the nominal board TDP. The power supply demands when throw in a 'big budget' power CPU are going to be more problematical if stick to the historical power supply levels.
forums.autodesk.com
