Danny PattersonNov 15, 2024
Paul Hudson
I just watched a comparison of M4 chips against dedicated NVIDIA GPUs, and it's just astonishing – M4 Max on MBP came in second to a 4080 Super on a desktop PC. If you're not sure why that's incredible, you should see the size and weight of a 4080 😂 https://www.youtube.com/watch?v=0bZO1gbAc6Y
How good are the Apple M4 chips when rendering in Blender?

YouTube
Nov 13, 2024 at 9:10amWeb
Show threadMatteoNov 13, 2024
@twostraws it’s incredible how fast it is, but I think the chart would have made more sense in a linear comparison!
Show threadRobert AtkinsNov 13, 2024
@twostraws Huge difference between the M4 mini and the M4 Max.
Show threadFrancesco MariniNov 13, 2024
@ratkins @twostraws M4 Max has up to 40 GPU cores, the base M4 only 8–10, and it might as well be that the clocks are (slightly) different, too. So not unexpected ☺️
Show threadRobert AtkinsNov 13, 2024
@fmarini @twostraws Must be a meaningful clock difference, the M4 Max is way more than 4× the M4 mini on that graph—and you expect basically a linear speedup per core for “embarrassingly parallel” problems like graphics right? Remarkable in any case.
Show threadFrancesco MariniNov 13, 2024

@ratkins @twostraws M4 Max is 2:27 mins, so 207 seconds. M4 9:13, 553 seconds. So the speed up is 2.67.
If the M4 is the 8 GPU cores version, and the Max the 40, theoretical scaling should be a perfect 5. If the M4 is the 10 cores, and the Max the 32, scaling should be 3.2.

Don’t be fooled by the flashy but imprecise graph 😉 Would’ve been better if they graphed speed up vs base M4, but 🤷🏼‍♂️

Show threadFrancesco MariniNov 13, 2024

@ratkins @twostraws if anything, that benchmark shows that Cycles isn’t able to scale linearly with GPU cores with the AS chips.

Oh, and they should have clearly stated which M4 and M4 Max version they are benchmarking…

Show threadFrancesco MariniNov 13, 2024
@ratkins @twostraws I stand corrected, made a stupid error (🙄🤦🏻‍♂️) converting mm:ss to seconds for the Max, see @alejandrolanza answer to my post 😅
Show threadAlejandro LanzaNov 13, 2024
@fmarini @ratkins @twostraws 2:27 is 147 secs. Specs are in the video, 10 vs 32 cores. So a 3.76 increase in performance for 3.2 times the cores
Show threadFrancesco MariniNov 13, 2024

@alejandrolanza @ratkins @twostraws yes, my fault 🤦🏻‍♂️😅

Didn’t watch the video, tbh.

So, if it’s 32 vs 10 cores, and 3.76 speed up, it suggests the clocks are different, at the very least +17.5% for the Max, but I would guess a bit more than that, since perfectly linear speed up is almost never achieved.

That said, those numbers are really, really good. And comparing perf/watts would make them even (much) better 👍🏻

Show threadAlejandro LanzaNov 13, 2024
@fmarini @ratkins @twostraws I haven't seen anything about different core clock speeds, but Apple did make a big deal about increased mem bandwidth from vanilla to pro to max, so that might be at play here
Show threadCraig MorganNov 13, 2024
@twostraws a 4080 Super alone idles at 15W when its doing nothing (ie. Windows desktop) … the M4x models perf/watt make a mockery of it … power draw measure under load would be the most interesting metric