Blender Discussions

[jl5086]

I'm wondering about the actual usefulness of being able to use more than one GPU in Blender for rendering.  I have an RX5700 XT and an RX580 8GB in my system, my question is what benefit do I get from having the option to use both GPUs?  What I have found is that, when using ProRender, any time I render using anything other than just one GPU, the render speed drops dramatically.  Just like selecting the option to use the CPU + GPU: it renders at about 1/2 the speed of just the GPU.  If I select just both GPUs: a specific render goes from 18 seconds up to 31 seconds.   Also when I use Cycles and I choose either the 580 or 5700, individually, the render times are what I would expect, the 5700 being a bit faster than the 580.  If I choose both in cycles it does impact the speed, but not as drastically as in ProRender.   It's harder to compare because in ProRender, if I select just the 580 as the sole render device, it still uses only the 5700 (which has my monitor plugged into it), so it only actually uses the second GPU if I have BOTH of them set to be used, and of course then it's about 30% slower than just the 5700.  If I add the CPU into the mix, then it's about half as fast as that!  So if that is normal, which it may be, then what exactly is the benefit of being able to use any of them together, and to not be able to use a secondary GPU on it's own?  I assume it's a bit of a bottleneck communicating between the 2 GPUs and the CPU, so that may be slowing it down.  I can't see the speed being any benefit, maybe being able to use the extra memory in the additional GPUs?  Even then you have the out of core memory option to use also.  It just seems pretty useless to have your render be a third slower and also have both GPUs using 150w of power each, unless it's in some way to aid in multitasking render layer situations.

[mrqozy]

wow!

just noticed this!

this is a deal breaker.

with one gtx 1070 i render in 7 Seconds.

with dual gtx 1070 i render in 15 seconds!!

[jl5086]

Yeah, I am hoping someone with the smarts to know will chime in and give a real world explanation for what the benefits are.  Cycles seems to also be slower on 2 GPUs, but at least it would let me use whichever one I picked.

[mrqozy]

Nope, cycles is faster with 2 GPUs

i justed tested it

one GPU 42 seconds.

two GPUs 26 Seconds

[jl5086]

Well I'll have to try again later,  I was only trying the veiwport render with cycles, and I didn't notice any speed improvement with both gpus, maybe for some reason it makes a difference that you have cuda or optix,  or that your gpus are identical.  I will let you know what my results are after.

[arozghon]

Hello! performance dropdown with mGPU with only one reason: Adaptive sampling, now it deactivate with mGPU and CPU+GPU modes. In one of the future builds, this fix will be added to RPR. Thank you

[mrqozy]

i hope you guys make this happen.

im very happy with the viewport denoising.

finally blender has it thanks to amd

[arozghon]

thank you for the feedback! 

[bsavery]

With the latest build (2.1) adaptive sampling should be enabled in multi GPU.

[jl5086]

I was not even aware of the 2.1 release!  I will retest soon, Thank you.

[jl5086]

Version 2.1.1 of ProRender is only usable at all if I don't touch the render quality selection,  because it's right to the desktop, and like the past versions when it crashes, 9 out of 10 times it leaves it in a state where the whole computer freezes up for all but 2 seconds about every 10 seconds.  So if I am lucky I can get to shut it down to restart before it locks up completely.  If it just crashes without becoming barely usable, I still have to re-start the computer because if I try to reopen Blender again, Blender will just lock up and crash anyway as soon as it gets to the gray screen and starts loading the plugin.   If I can't get to the start menu and restart it I am forced to just press the reset button and force it to restart.  That makes it really frustrating to try and evaluate these versions, so after 4 or 5 times that happens I always have to just go thru re-installing an old version of ProRender without full spectrum.  I don't get any errors to report whenever this happens, so I kind of have no where to go with this,  The only Noticeable thing with the version 2.1 of RPR was that when switching to rendered mode (full), the "Starting" time takes WAY longer than before, even with an empty scene it would take 30-40 seconds.   I really wish I could make this work better, I have an all AMD system, with their latest products and I can't seem to get a handle on why a simple scene with almost nothing in it would crash so easily.  I will continue using old versions, and I guess using my 2nd GPU will have to wait.

[jl5086]

So with latest RPR 2.1.1 selecting both GPUs uses both GPUs.  Selecting only Primary GPU(w/display) uses primary only.  Selecting secondary GPU only, actually uses Primary GPU only.   So the only way to use my 2nd GPU with ProRender is to use it combined with my primary GPU, but that way gives the EXACT same render times as just my primary GPU alone.(yet they are both being utilized, I verified it)   Although Cycles WILL use each GPU independently if I choose, the performance is similar to RPR results, being that with the RX580 + RX5700 XT combined is about the same render time as the 5700 on it's own.  and the 580 on it's own is slightly slower.   Is this likely that splitting between the two GPUs is not going to be beneficial because the 580 is much slower, and is slowing the whole process down?