ratGPU OpenCL raytracing benchmark

[jogshy]

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

Yep, 773s sounds good for that ATI.

Although we optimized for ATIs, I'm not completely sure why but they perform a bit slower than NVIDIA cards with the current implementation.
Taking a look to ATI Stream's profiler I see a lot of SIMD registers being flushed for secondary rays ( which are very incoherent ). That, combined with the lack of DMA support in the ATI OpenCL SDK v2.2 and the incapacity to execute the SKA profiling Tool(my kernels just crashes it) probably makes them a bit slower compared with NVIDIA cards.

Anybody tested a 68XX, pls? I've heard they are using 4D VLWI instead of 5-VLWI so less registers will be flushed and the performance could be a bit better.

Btw, I think for the next version we're gonna low a bit the benchmark's default settings to complete the test sightly faster ( and we'll try to perform several optimizations we have in mind also ).

[STEvil]

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




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 ...

Yeah, I used those registry changes to pass the #4 test. I set the GPU watchdog timeout to 200.

[jogshy]

Yeah, I used those registry changes to pass the #4 test. I set the GPU watchdog timeout to 200.

Nice, I'll tweak the installer for the next version based on that. Thanks !

[jogshy]

I've posted the 0.4.4 with several optimizations and some minor bugs corrected.
Now it takes around 88s for a GTX460 instead of 600s.

[STEvil]

http://www.ratgpu.com/verify.aspx?k=...C4360D167520C5

43.547 on the GTX480 @ 800/1600/2000 WinXP-64 with 0.4.4

[Arctucas]

4.4@875/2125

[Arctucas]

4.4@925/2200

[tet5uo]

Wow this is faster now, lol.

[jogshy]

Nice results !

By the way, caution with the Catalyst 10.10 in case you use an ATI card. Seems they're bugged. You need to use the old Catalyst 10.9 with the OpenCL SDK v2.2.

I'm investigating other optimizations and trying to port it to OpenSUSE 11.3 and Fedora currently.

[Chumbucket843]

Yep, 773s sounds good for that ATI.

Although we optimized for ATIs, I'm not completely sure why but they perform a bit slower than NVIDIA cards with the current implementation.
Taking a look to ATI Stream's profiler I see a lot of SIMD registers being flushed for secondary rays ( which are very incoherent ). That, combined with the lack of DMA support in the ATI OpenCL SDK v2.2 and the incapacity to execute the SKA profiling Tool(my kernels just crashes it) probably makes them a bit slower compared with NVIDIA cards.

a few questions:

it looks like you are using path tracing by the noise and GI quality (ambient occlusion seems very accurate.) but i see familiar pixel aliasing which leads me to believe this is a forwards ray tracer. which is it?

have you tried to improve coherency through sorting? you really need to do that with gpu's because it works well with data parallelism where as a mostly incoherent ray tracer works well with task parallelism.

if you are using forwards ray tracing are you using ray casting + primary rays or rasterizing like a hybrid engine would?

[jogshy]

it looks like you are using path tracing by the noise and GI quality (ambient occlusion seems very accurate.) but i see familiar pixel aliasing which leads me to believe this is a forwards ray tracer. which is it?

It's a path tracer but I re-utilise the primary rays's data to optimise a bit. I still need to implement the AA, yep.

have you tried to improve coherency through sorting?

Yep, I actually group the rays but some secondary rays are still very incoherent so the packets are flushed almost empty.

[Chumbucket843]

It's a path tracer but I re-utilise the primary rays's data to optimise a bit. I still need to implement the AA, yep.

i dont have much experience with path tracing but for the AA you could do something interesting such as using a gaussian filter or perhaps MLAA.

Yep, I actually group the rays but some secondary rays are still very incoherent so the packets are flushed almost empty.

damn. i have never optimized a ray tracer that much, mostly just simple things. i am sure you can find some good papers on improving coherency with gpu's although i dont know if they would be relevant to path tracing.

[jogshy]

I've uploaded the OpenSUSE 11.3 port. Here's an screenshot:



Fedora and FreeBSD will be my next victims
I would like also to target MacOSX Leopard but I lack a mac

[karbonkid]

Hello sir,

This is a very nice app. Are you planning a Blender exporter? What about real-time scene navigation like in Octane? CPU+GPU rendering like Lux? And finally, will it remain free or are you planning to charge eventually?

[jogshy]

Are you planning a Blender exporter?

Currently I'm working only in Maya and XSI exporters but, after that, it's possible, yep.

What about real-time scene navigation like in Octane?

Well, I have to do a lot of things before that but maybe possible, yep.

CPU+GPU rendering like Lux?

You can actually enable hybrid rendering. However, the CPU is much more slower than the GPU so "the chain is as strong as its weakest link", almost for small images. For large images you can get a benefit but for small ones you'll get worse results because the last thread usually must wait for the CPU to complete the image.

And finally, will it remain free or are you planning to charge eventually?

Well, I currently do this as a hobby.

[Spectrobozo]

very interesting software

here the result of the benchmark on a 9600gt under windows xp

[karbonkid]

Well, I currently do this as a hobby.

Thanks for the answers sir.

Have you considered open-sourcing it?

[NapalmV5]

4.4@925/2200

900/2200

[jogshy]

900/2200

Btw, FYI, I think the graphic drivers reserve certain Mhz/VRAM Mbs for the Windows's GUI, that explains the discrepancy between GPUz and the quantities that ratGPU reports.

[Arctucas]

900/2200

900/2200

[Arctucas]

930/2200

[Arctucas]

950/2200

[jogshy]

950/2200

Wow! 50Mhz for 1Ghz!

Btw, I've not test it but there's a trick you could use to get better score in SLI/Crossfire:

1. Use other card as primary adapter ( to plug the monitor there... for instance, a GT240 ). UNcheck the "Use this device" to avoid to use this card for the benchmark. This card will be dedicated to paint the Windows's GUI exclusively.

2. Check-in the "use this device" for the two GTX460s only.

In that way the GTX460s won't spend time painting the Windows's GUI, they will be 100% dedicated to the benchmark... so the score should be sightly better.

[-Acid-]

62.154 sec with stock 470 gtx

[tet5uo]

I ran it on the rig I just put together with 3 way SLI of gtx285's. 39 seconds.

Trying to find where I saved the screen :P