NVIDIA OpenCL Linux Performance Benchmarks

[phoronix]

Phoronix: NVIDIA OpenCL Linux Performance Benchmarks

Earlier this week I shared some AMD Catalyst OpenCL benchmarks showing the performance gains made by AMD's proprietary Radeon graphics driver this year in bettering the GPGPU results. In this article is similar testing on the NVIDIA side with their binary driver...

http://www.phoronix.com/vr.php?view=MTI1ODU

[oleid]

Phoronix: NVIDIA OpenCL Linux Performance Benchmarks

Earlier this week I shared some AMD Catalyst OpenCL benchmarks showing the performance gains made by AMD's proprietary Radeon graphics driver this year in bettering the GPGPU results. In this article is similar testing on the NVIDIA side with their binary driver...

http://www.phoronix.com/vr.php?view=MTI1ODU

It seems as if the Radeon 6950 -- which costs around 200 EUR is a) much faster and b) also much faster per EUR than this NVIDIA card -- which costs around 140 EUR.

AMD Score / NVIDIA Score
Room: ~ 2
Sala: ~ 2.1
Ball: ~ 3.3

And the price of the AMD / NVIDA card is only 1.4. So even if you scale the result with the price of the card, the AMD hardware wins...

(AMD Score/EUR) / (NVIDIA Score/ EUR)
Room: ~ 1.4
Sala: ~ 1.5
Ball: ~ 2.3

So I'm wondering... is it NVIDIA's OpenCL implementation which is slow, is it only this specific Benchmark in which AMD wins or is AMD hardware generally better suited for computational workload?

[Nille]

In my practical knowledge AMD is faster but Nvidia has an much much better Developer Support and Documentation. I have nothing from am cheaper card if i have to invest my time to get things working well.

[oleid]

In my practical knowledge AMD is faster but Nvidia has an much much better Developer Support and Documentation. I have nothing from am cheaper card if i have to invest my time to get things working well.

Developer Support and Documentation? Can't you use any OpenCL documentation out there?

And as an end user... when not writing programs by yourself, the amount of documentation and developer support shouldn't matter...

[Nille]

Can't you use any OpenCL documentation out there?

The Hardware is to different. Sure you can write simple openCL code but then expect that he is slow. You has to optimize it for each Hardware vendor. For this Optimizations you need the Documents and Support from the Vendor.

And as an end user... when not writing programs by yourself, the amount of documentation and developer support shouldn't matter...

Indirect yes. But if no one develop the Software on AMD Hardware and test maybe only if it works, the user has nothing from the fast hardware if he is not well used. As an example. Try to get an OpenGL SDK from AMD that is not completely outdated. Or an bit simpler try to get the OpenGL headers from AMD. The Radeon SDK is extremely outdated ( 2005 ).

Please Note that the khronos group refer to the hardware vendors for this.

[semhustej]

It seems as if the Radeon 6950 -- which costs around 200 EUR is a) much faster and b) also much faster per EUR than this NVIDIA card -- which costs around 140 EUR.

AMD Score / NVIDIA Score
Room: ~ 2
Sala: ~ 2.1
Ball: ~ 3.3

And the price of the AMD / NVIDA card is only 1.4. So even if you scale the result with the price of the card, the AMD hardware wins...

(AMD Score/EUR) / (NVIDIA Score/ EUR)
Room: ~ 1.4
Sala: ~ 1.5
Ball: ~ 2.3

So I'm wondering... is it NVIDIA's OpenCL implementation which is slow, is it only this specific Benchmark in which AMD wins or is AMD hardware generally better suited for computational workload?

I think the developer of LuxMark uses AMD card, so the code is more optimized for AMD cards. I guess with OpenCL, architecture optimisations can make huge difference.

[Nille]

I think the developer of LuxMark uses AMD card, so the code is more optimized for AMD cards. I guess with OpenCL, architecture optimisations can make huge difference.

So is it. OpenGL/OpenCL is Hardware dependent but Platform independent

[rudl]

That benchmark doesn't make sense without the 275.21 driver, those were 2 times faster in the complex luxmark benchmark , and a lot faster overall !!!!!!!

Yeah Fuck you nvidia!