ratGPU OpenCL raytracing benchmark

[jogshy]

Hi!

We've created a free program to render images using raytracing and the GPU and to benchmark the OpenCL's performance. It's called ratGPU and you can get it at www.ratgpu.com ( which redirects temporally to www.xnormal.net/ratGPU ).

There is small benchmark mode with web verification. If you could test it a bit it would be really appreciated.

Usage:

1. Download and install the program ( www.ratgpu.com ). You should reset after installing it ( see limitation #3 to know why ).

2. Execute the StandAloneRenderer.exe ( easily from the "start" menu S.Org->ratGPU->StandAloneRenderer ). Notice if you use a 64bits OS like Windows 7 x64 you should execute the "Standalone renderer (64bits)" or, due to a current limitation of the Forceware drivers, the program will fail. So, you cannot execute the 32bits version currently in a 64bits OS. Use the x86 version in a 32bits OS and the x64 version in a 64bits OS as is natural.

3. Press the "benchmark" button on the top of the toolbar and wait until it finishes.



You can see the progress of the current test or to cancel it with the button.



4. Copy the verification URL link and save it if you need to validate the result after ( for example, for HW-bot ).




Limitations/problems:
0. As it uses the GPU, your monitor or Windows's UI can be very unresponsive.

1. Old drivers ( Forceware 197 ) might have problems. The FW 25X are problematic too. You should use the FW 260 Beta atm.

2. If you use an ATI card, you must download and install their OpenCL SDK runtime ( because Catalyst drivers don't include it... don't ask me why... )
Download it at http://developer.amd.com/gpu/atistre...s/default.aspx
Note: Only Radeon 5XXX are fully supported, 4XXX series lack the required functionality.

3. If you use Windows XP, a thing called "GPU watchdog" could pop resetting the graphics drivers and aborting the program execution. That's a thing Microsoft added in the 90s to reset the driver in case a program would be stressing the GPU for more than 5 seconds. Of course, that was before GPGPU existed because modern OSs like Vista/Windows 7 include a mechanism to avoid that ( I do it automatically in the installer, but for WinXP simply cannot be disabled ).
To solve that problem I suggest you to get a second GPU and disable the primary card ( the one connected to the monitor ) using the "config" tab and, then, uncheck the "use this device" checkbox.
Again, nor Vista/7 requires that because the watchdog is disabled automatically.


My score with a reference Gainward GTX460 (192bits), W7 x64, FW260b and an i7 920 is 598s-601s.
Here is a screenshot showing several things:



Btw, we're porting it to Ubuntu 10.10 too.

Thanks!

[Sn0wm@n]

no ati 4290 love ????

[jogshy]

no ati 4290 love ????

I'm afraid not. ATI said the Radeon 4XXX series have too many OpenCL limitations:

http://www.xbitlabs.com/news/video/d...rformance.html

Only the 5XXX ones are supported because the 4XXX ones cannot use more than one hardware buffer and I'm using more than ten
For NVIDIAs, any card equal or above than a G80 should work, but the G80's performance under OpenCL is seriously limited. I recommend to use a GT200 or Fermi or the test will take forever.

Btw, you can also use the CPU in case no GPU could be used... but it will be much much much slower. Just select your CPU in the "device" combobox.
Also, you can load one of the example scenes located in the [installdir]\StandAloneRenderer\scenes folder, go to the "renderer" tab, select "Path-tracing" and press render. In that way only that scene will be rendered and you can also make the image smaller if you need a faster test.

ps: edited/updated "limitations"

[Sn0wm@n]

thanks for the info man

[STEvil]

Gave it a go and couldnt quite get it to finish..

CPU renderer crashed out
GPU renderer got to the 4th or 5th scene (?) then a pop-up came up saying "ratGPU cannot render the tile" or something to that effect.

system is:

2x E5450 @ 3.5ghz (LGA771, 4cores per CPU)
8gb ram
GTX480 with 260.63 drivers, 797/1599/1998 clocks
WinXP-64

[Arctucas]

Mine-SLi GTX460 @ 860/2100.

[jogshy]

GPU renderer got to the 4th or 5th scene (?) then a pop-up came up saying "ratGPU cannot render the tile" or something to that effect.

system is:
....
WinXP-64

For simple scenes like the ambient occlusion one or the balls probably the watchdog didn't pop, but the last scene is very complex so if a pass took more than 5 secs to execute then the watchdog probably popped causing the error.
For winXP you might need to use a second card and to disable the primary adapter or the GPGPU watchdog will pop. I'm afraid WinXP is not really prepared to perform GPGPU with only one card 8( With Vista/7 should work better.


You mention the CPU device crashes... have you tried it without OC, pls? The CPU device test stresses all the cores so it might be due to OC, but I'll revise the code searching for bugs just in case.
Btw, are you using only the CPU device or in combination with the GPU ones?

Mine-SLi GTX460 @ 860/2100.

Nice SLI scaling! ( > 2x my score! ). Are you using 192bits cards or 256bits ones?


Btw, you can use BOTH cpu+gpu hybrid rendering ( play a bit with the "use this device" option ), but sometimes the GPU has to wait the CPU so the result might be worse than using only the GPU. If you disable the GPU device and you use only the CPU you'll notice the GPU is much much much faster. I recommend to check in only the GPU devices to render fast and only the CPU to perform a CPU OC stress test.
For instance, compare the 600s of the GPU-only renderer vs the 2000s of the CPU-only:



CPU+GPU together did 721s, worse than GPU-only because my CPU is much slower than a Fermi card so it only slowed down the GPU ( a chain is only as strong as its weakest link ).

[Arctucas]

@jogshy,

Thanks.
The cards are eVGA SuperClocked External Exhaust 1GB versions-01G-P3-1373-AR, so, 256-bit.

Here is a run @900/2100:

[STEvil]

For simple scenes like the ambient occlusion one or the balls probably the watchdog didn't pop, but the last scene is very complex so if a pass took more than 5 secs to execute then the watchdog probably popped causing the error.
For winXP you might need to use a second card and to disable the primary adapter or the GPGPU watchdog will pop. I'm afraid WinXP is not really prepared to perform GPGPU with only one card 8( With Vista/7 should work better.


You mention the CPU device crashes... have you tried it without OC, pls? The CPU device test stresses all the cores so it might be due to OC, but I'll revise the code searching for bugs just in case.
Btw, are you using only the CPU device or in combination with the GPU ones?


Nice SLI scaling! ( > 2x my score! ). Are you using 192bits cards or 256bits ones?


Btw, you can use BOTH cpu+gpu hybrid rendering ( play a bit with the "use this device" option ), but sometimes the GPU has to wait the CPU so the result might be worse than using only the GPU. If you disable the GPU device and you use only the CPU you'll notice the GPU is much much much faster. I recommend to check in only the GPU devices to render fast and only the CPU to perform a CPU OC stress test.
For instance, compare the 600s of the GPU-only renderer vs the 2000s of the CPU-only:



CPU+GPU together did 721s, worse than GPU-only because my CPU is much slower than a Fermi card so it only slowed down the GPU ( a chain is only as strong as its weakest link ).

Ah, the hybrid rendering is what made it crash. I didnt realize you could have multiple devices selected..

As to the GPU tests, they for sure took more than 5 seconds and didnt cause the error until that ~4th test (final test?).

Giving these a try to see if they help:

http://msdn.microsoft.com/en-us/library/ff553890.aspx
http://msdn.microsoft.com/en-us/library/ff553893.aspx


edit..

oh wow.. that was fast lol. 282.2

http://www.ratgpu.com/verify.aspx?k=...EB1CE716751B0C

[jogshy]

Giving these a try to see if they help:
http://msdn.microsoft.com/en-us/library/ff553890.aspx
http://msdn.microsoft.com/en-us/library/ff553893.aspx

Interresting, thx! I'll try tomorrow.

edit..
oh wow.. that was fast lol. 282.2

Now you own the WR

Btw, I've just uploaded the 0.4.3 with some minor bugs corrected and Ubuntu 10(.04/.10) support:

[Arctucas]

Using the 4.3 version, @925/2200:

[tet5uo]

Ran it on 3 480's and a 450

Clocks on the 480's were 867/1734/1988 and the 450 was 930/1860/1900
i7 980x at 4.4gHz.

[jogshy]

Ran it on 3 480's and a 450

Wow, nice!

Btw, caution if you mix a card much slower than the rest. Perhaps 3X 480 could be faster because, with a small image like the one used for the benchmark (512x512) , the "chain is as strong as its weakest link"

[tet5uo]

Wow, nice!

Btw, caution if you mix a card much slower than the rest. Perhaps 3X 480 could be faster because, with a small image like the one used for the benchmark (512x512) , the "chain is as strong as its weak link"

Cool, yeah I noticed that 4th thread working alot slower than the rest

I'll give it a shot with just the 480's now....

Yep faster

[jogshy]

I'll give it a shot with just the 480's now....
Yep faster

Wooow! I wonder what 4x 480s could do!

Btw, anybody can test any ATI card, pls? Although in all my experiments ATI cards were much slower, even when I spent a lot of time optimizing for it. I think the incoherent ray tracing's secondary rays are flushing a lot the Radeon's VLWI SIMD registers and the lack of DMA support is slowing down the PCI express transfers and kernel calls
Btw, remind to use ratGPU with an ATI card you must:

1. Get a Radeon 5XXX or one of those new 6870s . Radeon 4XXX are not supported ( because they lack the required OpenCL functionality ).
2. Download and install the ATI OpenCL SDK v2.2 at http://developer.amd.com/gpu/atistre...s/default.aspx
3. Use Catalyst 10.9 or above.

[Manicdan]

does this work with only CPUs too? and if so, about how long would that take on faster quad/hexies?

[tet5uo]

Wooow! I wonder what 4x 480s could do!

.

They could make my power bills even more obscene

I've already seen this system draw 1.38 kW from the wall with max GPU load in furmark

[jogshy]

does this work with only CPUs too? and if so, about how long would that take on faster quad/hexies?

Yep, ratGPU can perform CPU-only renderers ( also GPU-only or GPU+CPU-mixed-mode renders aka "hybrid raytracing" ).

To use only the CPU just select your GPUs with the Config->Device combo box and disable them with the "Use this device" check box. Once all your GPU devices have the "Use this device" UNchecked just select the CPU device and check in its "Use this device".

If you render with only the CPU device you'll notice the GPU is really good performing ray tracing. My i7 920 OCd to 4Ghz takes 2000s vs the 600s of my GTX460, so it's about 3.5X slower than my GPU.

[STEvil]

Assuming 100% scaling, dual westmere's @ 4ghz should take 667s.

Will have to test on my SR-2 when its up in a week or two hopefully, if someone doesnt beat me to it.

[tet5uo]

does this work with only CPUs too? and if so, about how long would that take on faster quad/hexies?

Here's a CPU only run on my 980x @ 4.45 w/ 12gb of ddr3 2000mHz cas10

[jogshy]

Here's a CPU only run on my 980x @ 4.45 w/ 12gb of ddr3 2000mHz cas10

Ooh nice! So a dual socket will use about 600s, yep. Funny to see 2x 1000$ CPUs + the SR-2 MB catching a 160$ card

Giving these a try to see if they help:

http://msdn.microsoft.com/en-us/library/ff553890.aspx
http://msdn.microsoft.com/en-us/library/ff553893.aspx

By the way, did you use that to make WinXP to pass the test #4, pls?
For Vista/7 is easy and I do it in the .exe installer but a graphics driver engineer told me it was virtually impossible to disable the watchdog for WinXP ...

[tet5uo]

Hehe yeah, here's my 130$ after instant rebate card kickin my 1k$ cpu's butt...

[tet5uo]

Ops, just realized I had steam open downloading a game while doing that CPU run.. might have shaved a few secs off without that running hehe

[tet5uo]

I'd test an ATI card for ya, but the only one I have here is in the closet, an old AGP x800XT :P

[p2501]

Is this reasonable for a HD5850 @ 900/1253?!

Catalyst 10.8:


Catalyst 10.9:


Not much of an improvement between both drivers IMO.