LuxrenderGPU

[fredmj]

I think it will still be a lot cheaper to go for GeForce cards for small installations

Yep, I think so. Nevertheless, I would be (VERY) curious to see the result of a rendering with luxrenderGPU on a TESLA machine...
Is there any rich man in the audience?

"NVIDIA NV100 (or Fermi) is less powerful than GeForce GTX 285?"
http://www.geeks3d.com/20091117/opencl-and-gpu-computing-latest-news-2/
C2050 1.040 TFLOPS
C2070 1.26 TFLOPS
GTX 285 1.063 TFLOPS

EDIT
GTX 295 1.788 TFLOPS
ATI 5970 4.64 TFLOPS

According to the same source :

C2070 stands a bit better, 630 GFLOPS of Dual-Precision [...]

and "dual precisions" mean operations over 64bits precisions, which is, I guess, the normality in the 3D world. No?
I know that is a bit out of the subject of this post and That's my last comment about that, but ...
For example, Is the MSI mother card (with her 7 PCI-E slots) plugged with C2070 could really explode the performance of LuxrenderGPU?
I mean 7X600GFLOPS=5,4TFLOPS, and the (pseudo) real time rendering come true...

[jensverwiebe]

and "dual precisions" mean operations over 64bits precisions, which is, I guess, the normality in the 3D world. No?
.

Nope..not atm.. most serious developer say the could live very well with singleprecision, but there are exeptions for some areas.


Jens

[fredmj]

OK I can beleive that.
But I always thought that the 3D applications was the only one who really take advantage of the double precisions.

[Chiaroscuro]

[...]But I always thought that the 3D applications was the only one who really take advantage of the double precisions.

Scientific applications... a Fermi target. For 3D CGI applications, not so much.

[TerabyteST]

Can anyone build this from SVN for windows? Would be great!

[vertigo262]

This is all great news. But I don't think I can wait for v0.8

Let's get coding!!!!!!!

I'd help if I had any coding talent

[KyungSoo]

Thank you for the Great Advance.

And, I tried it with 3 machines and the results were as follows;

1. 2 CPU + 4 GPU machine: about 500K samples/sec
2. 1 CPU + 8 GPU machine: about 250K samples/sec
3. 1 CPU + 1 GPU machine: about 250K samples/sec

The performance was not depanding on the number of GPU,
it was depanding on the number of CPU, by now.

Conclusion: There is a bottle neck on rayBuffer feeding.

I am looking forward to the day, getting 4M samples/sec with my 8 GPU machine.

Reagrds,

[Dade]

Thank you for the Great Advance.

And, I tried it with 3 machines and the result was as follows;

1. 2 CPU + 4 GPU machine: about 500K samples/sec
2. 1 CPU + 8 GPU machine: about 250K samples/sec
3. 1 CPU + 1 GPU machine: about 250K samples/sec

The performance was not depands on the number of GPU,
it was depands on the number of CPU, by now.

Conclusion: There is a bottle neck on rayBuffer feeding.

Oh, for sure, keeping the GPUs well feed is the major problem at the moment but have you modified the code for your hardware configuration ? It is all hardcoded (i.e. number of native threads, number of threads used per GPU, etc.) at the moment

I am looking forward to the day, getting 4M samples/sec with my 8 GPU machine.

I would be very curious to see a screenshot the latest version of SmallLuxGPU on your 8xGPUs hardware, can you give it a try ? viewtopic.php?f=34&t=3458

Can you post also some detail about your hardware (i.e. kind of CPU and GPU) ?

[KyungSoo]

I tried the test, with luxRender, openCL version 0.7.
And I did not modify it at all.

And I tried SmallluxGPU v1.2 too, now.
It performed 176K, 189K, 178ᆞK, 173K,ᆞ 186K, 179K, 186K, 189K for each GTX 295 ᆞGPUs,
and load percentage was around 40s for each GPUs, and the total average. samples/sec was 2,558K.
But, I did not attach the screen dump, since I do not know how to do it.

And SmallluxGPU v1.3Beta1 test was not successful.
I met following messages.
[Device::GeForce GTX 295] RayIntersection thread started
[Device::GeForce GTX 295] RayIntersection ERROR: clEnqueueWriteBuffer(-5)

As for as I know SmallluxGPU is a small test bed of luxRender.
But the results are too different between two environments.

[fredmj]

8 GPU's machine?
Just because I am curious, could you give me details about your configuration?
I mean, Is it two 4-Way sli machines, or 4 machines each double SLI?