I have a few 2D works that I tested, and in M2Max (30GPU) I got about 6x the duration compared with RTX 2070 (PC). I returned the M2Max and kept working on the M1 16GB MacBook Air. Oh an M3 Pro locked with RT ON is 2x the speed of the same PC, meaning is much faster than M2Max with almost half the price.
Yes I saw ā admin is like: is this the 2% of population?!! Go to your 10k thread (now closed due to limitation).
This is interesting. The results Iāve seen so far have suggested that in Cycles roughly M3 Pro = M2 Max. If in some projects the M3 Pro is indeed 2x faster than the M2 Max then this is very good news - the M3 Max could be a monster.
So the metal RT is on by default? Is possible to confirm that the M3 is running RT in the graph above? Or are the gains purely from the node and architecture advancement?
Yes, the MetalRT is enabled by default for M3 Macs.
There are architectural improvements in the M3 processors, but not so much that without hardware raytracing you get 2.5x better results (the MBP M3 Pro tested had only 14 GPU cores and the MBP M2 Pro had 16 GPU cores). The M3 Max has 40 GPU cores in the full version, so, if Iām not mistaken, there would be about 20 seconds with linear scaling for this demo scene.
This thing is going to be dirty. Canāt wait!
The day will come when the Nvidia fanboys will have to say sorry to us.
ā¦ with up to 0,25 TB GPU memory?
So net 248 GB āVRAMā
Still 8 Apple base GB for the rest ā¦
I noticed that 4.1 with metalrt on performs same as 3.6 with metal rt. On M3pro.
(I tested this on blender classroom)
So I guess more performance gain is to be expected from further optimization? On my own production scene M3pro was 50% of 4070 in octane rendering. Or somewhere around a 2070. And this is definitely without optimization for hardware ray tracing.
On my desktop I usually prefer cycles in terms of performance but on mac it looks octane is somehow doing better.
And one more observation, in unreal the old school cascaded shadow maps are faster to render than virtual shadow maps on a dense scenery with lot of nanite meshes.
Just curious, but have the late 4.0 changes like Metal RT been pushed to 4.1 yet?
No idea, I just enable metal rt and full optimization in blender system setting and donāt see a difference in performance. Actually I think 3.6 was a little faster for some reason.
Or is there supposed to be a separate hwrt setting? I havenāt noticed anything.
Thatās why I was asking about 4.0 compared to 4.1.
4.0 is where the HWRT was turned on last week. Just wondering if that has been pushed to 4.1 yet? And why your 4.1 version maybe performing just like 3.6.
nvidia has very expensive cards with a lot more than the 4090ās 24gb vram though. price and vram competition will be good for nvidia users too. it could be ~10000$ mac studio vs. ~10000$ nvidia card then.
Steam Deck uses unified memory too and it just runs linux so thatās probably a more apt comparison than the Xbox or PS5 since both use GPDDR vs LPDDR. So itās very possible to do a unified memory system, they already exist.
The difference being, that the Steam Deck GPU can only access up to 8GB of the memory.
Yep. Iām assuming thatās an arbitrary cap that Valve could increase if need be.
Can you explain what you made here?
Could you be more specific?
This probably means that MetalRT version 3.6 supports hardware raytracing. This is not surprising to me - I wrote some time ago that the M3 would have HWRT support from day one. Except that MetalRT is only enabled by default from v.4 onwards.
There is a second explanation - you got the M2 Pro in the shipment instead of the M3 Pro
Ideally youād do tests (e.g. Classroom and Party Tug) of Cycles with MetalRT on and off. It would be great if you posted the results.