Will AMD FirePro Graphics cards work ok with VrayRT?

[vlado]

OK, So where can I go to get that version?

I've been sent to 'nightlies.chaosgroup...' by your collegues twice now but I see no file name like that..

It would be best to wait for a couple of weeks when we will have an open beta of the next service pack.

Best regards,
Vlado

[PGD]

Digging this old thread up again, but:

I just upgraded to V-Ray 3.30.01, and at first WOW - it worked. I had a couple of issues where the Active render would update for anything BUT geometry changes. So I could change the light position or Multiplier and the render changed, rotate the tea pot and the render updated, but ignored the moved tea pot - which seemed odd?

Shortly after this issue did resolve itself by closing the frame buffer window and trying again.

Then I ran a working scene..

So all I've done is open a previous scene of a building and made these changes to the Max file:

Assign render, change Activeshade to RT 3.30
Then change Target from Production to Active shade mode (which then loses the frame size settings, why?)
Change render engine type to OpenCL
Hit render, see some errors on the render, sub pixel mapping etc then get the message ‘Out of resources’

So - this doesn't seem to be working yet

[savage309]

Some errors on the render tab does not sound too good.
Can you re-open Max and try again or share the scene, so we can test it ?

Thanks.

p.s. keep in mind that if the GPU memory can't fit the scene, you will most likely get the same out-of-resources error.

[PGD]

I'm experimenting with a 364 KB 3DS scene with a few tea pots and a couple of plane lights - yet I'm getting the 'Out of resources' message. Anything I can do?

[savage309]

Wrong post, just ignore.

[savage309]

I think I've missed to ask - what is your driver and Windows version ?

[PGD]

Win 7 Pro, ATI driver (14.502.1019.0)

[savage309]

Win 7 Pro, ATI driver (14.502.1019.0)

It is our fault to not notify about the requirements for OpenCL rendering on AMD.
You will need driver 15+. For gaming cards it is available for Windows 7,8 and 10. For pro cards it is available only for Windows 10.

I will make sure to find appropriate place in the forums/docs/site to make those requirements easier to find from our users. Sorry for that.

[PGD]

Sooo...

We've 'upgraded' to Win10 to allow us to run the new graphics drivers.. Then ugdraded the graphic drivers, Current version is: 15.201.2401.0000 – Installed 20/01/16

As a reminder the graphics cards are:
AMD FirePro W9100 - 16GB GDDR5 Workstation Graphics Card, and a
AMD FirePro W5100 - 4GB GDDR5 Workstation Graphics Card


I've picked the GPU Benchmark scene up from this thread and tried to give it a go. (http://forums.chaosgroup.com/showthr...400-test-scene) I attach a screenshot of the setup in the hope that helps. The render looks preety weird, doesn't it?

Here's the render message:

warning: Building OpenCL trace program for device AMD Accelerated Parallel Processing_Hawaii. That build is made only the first time you run V-Ray on a particular hardware, but it might take a while.
warning: Failed to compile OpenCL kernels, falling back to CPU code.
error: buildProgram() failed for device 1 (-11)!
warning: initDevices() failed. Please check that your system has compatible GPU devices installed.
error: 1 error(s), 4 warning(s)


If I switch to CPU mode the render doesn't look weird like that..

On render, with OpenCL mode selected I notice the CPU goes up to 99% in the Task Manager..

What do you think?

[savage309]

Is that V-Ray 3.30.03 ? We managed to break in there. Sorry for that, try with 3.30.04.

[bibi5000]

But the driver 15.201.2401-whql-firepro-retail is also available for win 7. I am currently testing this driver with Vray 3.30.04 on my FirePro w8100.
Runs fine for now, but did not with Vray 3.30.03.

[savage309]

Yes, it was released recently. Let us know if there are any issues, we could definitely use some feedback for the RT GPU OpenCL running on AMD GPUs.

[bibi5000]

Will do.

I grabbed the GPU benchmark scene and did some renderings, comparing rendering speed of the GPU and the CPU under OpenCL, with coherent tracing on and off.
I had to stop the rendering manually as soon as 527 paths/pixel where reached, so this benchmark is not as accurate as you would wish.

Just to be sure, I did a rendering in CPU mode which I cancelled after 10 minutes and 379 paths/pixel. But it looks quite different in terms of textures and camera FOV, so this seems to be no real comparison.

It struck me as odd that the CPU under OpenCL is faster than the GPU if coherent tracing is turned off. Can you confirm this?
And while combining CPU and GPU really helps, they do not seem to scale linear, meaning their speed does not add up.
I am quite disappointed that GPU and CPU show nearly the same rendering speed, but I would like to see if other benchmarks with FirePro cards show the same results.

Benchmark_comparison.pdf

[savage309]

Thanks, these type of things are the helpful feedback I was thinking about.

Just to be sure, I did a rendering in CPU mode which I cancelled after 10 minutes and 379 paths/pixel. But it looks quite different in terms of textures and camera FOV, so this seems to be no real comparison.

Camera distortion is not supported on the GPU, that's why the FOV looks different. This however does bias the results so much, so the results you got are actually pretty accurate. As you can see, RT OpenCL running on the CPU is actually faster than RT CPU in this particular scene setup. Drawback is that RT OpenCL CPU has all the limitation the RT GPU render has.

It struck me as odd that the CPU under OpenCL is faster than the GPU if coherent tracing is turned off. Can you confirm this?
Benchmark_comparison.pdf

It is the case that not every GPU setup will be faster than any CPU setup. If you have fast and decent CPUs they can be faster than some GPUs. I have Fury X and i7 3960x, Fury X rendering for 224s vs 297s for i7 3960x. i7 2600 takes 478s.
Keep in mind that we don't use any of the GPUs pro features, so having pro GPU doesn't actually help much with the RT performance.

And while combining CPU and GPU really helps, they do not seem to scale linear, meaning their speed does not add up.

RT usually scales linearly, the thing is that the CPU gives tasks to the GPU and renders as well. There is a way to improve that, I just wanted to know if anybody is interested in such setups.

I am quite disappointed that GPU and CPU show nearly the same rendering speed, but I would like to see if other benchmarks with FirePro cards show the same results.

The results of the newer AMD GPUs is better and quite competitive in some cases. I believe we could work on optimizing our code for AMD GPUs further especially if there is enough of interest.
Keep in mind that this is the first (of lets hope, many) V-Ray versions to work at all with AMD GPUs. Hopefully it will only get better from here.

[bibi5000]

There is a way to improve that, I just wanted to know if anybody is interested in such setups.

Since it is the strong side of OpenCL to adress CPU and GPU likewise it seems like a good idea to use that as good as possible. I tried the Benchmark 2.0 with GPU and Coherent tracing on. Result: 12 min 15.6 s which puts the w8100 in to the expected range of a Quadro M6000 according to the Benchmark Spreadsheet. The same scene on GPU+CPU, Coherent Tracing off, finishes at 7min 59s.

Since most people will get a Quadro or Firepro for their deep color support (like me) or some certified driver stuff, it is pretty interesting to see that a mid range FirePro card is performing pretty well against a Quadro six times as expensive.

The results of the newer AMD GPUs is better and quite competitive in some cases. I believe we could work on optimizing our code for AMD GPUs further especially if there is enough of interest.
Keep in mind that this is the first (of lets hope, many) V-Ray versions to work at all with AMD GPUs. Hopefully it will only get better from here.

Sounds exciting