WOW! whats going on here? super fast renders 2.79 beta

So i downloaded the official blender version. a few weeks back but i also downloaded a beta 2.79 from 25th of dec 2017.
I dont know if its settings i’ve altered or just the way each build works… But on the official build if i click GPU render with auto tilesize it uses the GPU only as confirmed by my task manager and GPU utility, and it uses one 200x200ish tile at a time. If i use CPU it uses 16 threads and 32x32 tiles and cpu only.
HOWEVER. If i use the beta version and select GPU render, it renders with like 12 200x200 tiles at simultaneously uses 100% cpu and 100% GPU at the same time and eats through the render about 10 times faster. I was stumped as to why my scene was suddenly rendering snail’s pace with one tile until it struck me i opened the document and it had auto opened with the official blender version not the beta.

am i missing something obvious? does everyone already know about this?

I believe that hybrid cpu/gpu rendering is the big difference coming but I must admit, I thought that was in V8.0

It’s faster, but not 10x faster, sound like something is wrong with the first render.

You’ll want to reduce tilesize in recent builds. Large tiles no longer have a big performance boost on gpu. The CPU does still take a performance hit from large tiles, so set your tiles to 16x16 for CPU+GPU rendering for the best speed. (32x32 if you are denoising)

I just tried on Windows OpenCL rendering the usual Mike-pan 2 BMW scene with my RX480. Latest Blender 2.79 version from Buildbot, latest AMD display drivers (Adrenalin).

32x32 tiles = 1:35:55
240x180 tiles = 1:03:71

Yea i was just going off of initial readings on the estimated render time left, but in reality it’s not likely to be that 10X, but it sure is faster.

yes even without reading your post i figured that and brought it right down to 75x75 and saw a phenomenal increase even further… Never imagined the GPU +CPU would like 16x16 though, thanks i will try it.

I think it will depend on your gpu brand. My dual RX 480 got the best performance at 64x64 with cpu+gpu. If I go cpu only, I get the best at 16x16 or 32x32. If I go gpu only, I get best performance at 256x256.

So here are some benchmarks and figures to let you know just how fast it is…
Classroom benchmark default settings, except i changed the tiles to 24X24 and have the setting where its using GPU +CPU…
4mins 46 seconds. According to some quick check on the internet that’s significantly faster than a 28 core 56 thread intel xeon…
i shall try BMW next.

System spec is AMD 1700X STOCK no OC… GTX 1070 stock no OC… 32GB DDR4. my cooling is weak and even slight overclocks on either leads to sudden pc shutdown, ah well i had to compromise somewhere as my PC build went way over budget anyway.

BMW gpu blend downloaded straight from the Blender website… 1min 37sec. So yea this is the new benchmark with two bmws one in foreground one in background. According to the internet i’ve beaten then 56 thread intel cpu from renderstreet again. also beaten their GPUS and anything else they could throw at it. Update. 16x16 tiles, 1 min 25. Next i will try a mild overclock to say 3.7GHz GPU and a tiny overclock on the GPU, and hope my PC does not shutdown.
Result was 1 min 23.81 sec. Only did 3.6GHZ on the CPU and a 50mhz oc on the GPU core.

Yeah, the CPU+GPU rendering is a significant boost. I’ve got a hexacore i7 that was just idling when I was rendering with my dual gpu setup. Now I get a 40% speed boost, no hardware upgrades required.

game changer.

Do you set something in Blender to get CPU+CPU or is it automatic? I have build blender-2.79.0-git.e428ea3-windows64 but don’t see a CPU+GPU setting

You didn’t download a “2.79 beta”, you grabbed a ‘master’ build which indeed has many improvements over 2.79 most of which will be in the official 2.80 release (but will never make it into 2.79a/b/c/etc)

Is this hybrid GPU/CPU thing a feature that is unique to blender?

Not to hyjack the thread… but I can’t find the CPU/GPU feature in the latest daily builds what build do I have to get to find it?

It’s there in the latest build available to download, you just need to go to: User preferences --> System then check the CPU box.


Will the CPU/GPU combined render feature be in the 2.79a release, or is it still considered to be in testing?

Wow, this is a really nice and unexpected belated Christmas present. Even with my old CPU a test scene of mine that takes 17m8s (GPU) now renders in 13M:12s (GPU+CPU). 4 minutes less!

Happy now :slight_smile:

There has to be something to do with CUDA 9.0 also. Because I went from GPU only rendering 33s down to 26.5s
When doing GPU+CPU 16px tiles I got 22s. So that’s pretty sweet upgrade, from the meeting notes it’s moving along, test builds this week. Can’t be so long until we have an ‘2.79 a’ release.

1) Blender 2.79 ‘a’ release

Release is going along nicely so far, you can track status here (for main repo) and here (for addons).

From the main repo,

CMake: support CUDA 9 toolkit, and automatically disable sm_2x binaries.]

Can anyone see that we will get CPU+GPU rendering in 2.79.a from the main repo notes?

— remove