NVIDIA OpenCL Linux Benchmarks

[phoronix]

Phoronix: NVIDIA OpenCL Linux Benchmarks

In testing of OpenBenchmarking.org and preparations for the release of Phoronix Test Suite 3.0-Iveland at the end of February from SCALE, a lot of benchmarks have been happening to test the various analytical features and other new capabilities of this open benchmarking platform. In fact, it is really an overwhelming amount of benchmarks; the power capacity in my office is maxed out as benchmark after benchmark and system after system there is all sorts of test scenarios being looked upon. The benchmarks coming out on Phoronix.com over the past two months have just been barely scratching the surface of what has been going into the OpenBenchmarking.org system. Recently a lot of OpenCL compute benchmarks were pumped in, and since we have only published a few OpenCL Linux benchmarks -- OpenCL on Linux vs. Mac OS X and OpenCL NVIDIA vs. ATI on Linux -- here's some more in this article.

http://www.phoronix.com/vr.php?view=15623

[bug77]

Something's wrong with the 9800GT. It should be about equal to the 8800GT.

Also, nice job posting "Even the low-end GeForce GT 220 graphics card is dramatically better with such workloads than a GeForce 9800GTX" right after the test where the 9800GTX scores almost 2500 and the GT220 "blows it away" with less than 550. Seriously, the GT220 beats the 9800GTX in only two tests. What did you mean?

[Shining Arcanine]

Something's wrong with the 9800GT. It should be about equal to the 8800GT.

Also, nice job posting "Even the low-end GeForce GT 220 graphics card is dramatically better with such workloads than a GeForce 9800GTX" right after the test where the 9800GTX scores almost 2500 and the GT220 "blows it away" with less than 550. Seriously, the GT220 beats the 9800GTX in only two tests. What did you mean?

Those were my thoughts too. This is not the first time I have seen something like this on Phoronix. It would seem that not much thought is being given to benchmark results, at least not in terms of any theoretical foundation.

http://www.nvidia.com/object/product...gt_220_us.html
http://www.nvidia.com/object/product...00_gtx_us.html

The GeForce 9800 GTX has more than twice the processing power and close to twice the bandwidth of the GeForce GT 220. If the GeForce 9800 GTX is losing to the GeForce GT 220, then either something is wrong or some sort of design innovation occurred that makes it faster. It would be nice if the review discussed this.

[andreas.h]

Interesting benchmark!

Are there any benchmark results comparing OpenCL and CUDA performance?


Cheers,
Andreas

[allquixotic]

Interesting benchmark!

Are there any benchmark results comparing OpenCL and CUDA performance?


Cheers,
Andreas

That's an apples and oranges type of exercise. Regardless of platform, hardware, etc., it is very difficult to compare the "performance" of different APIs. This is also why you very rarely see any serious discussion about the "performance" of Direct3D versus OpenGL.

In reality, what makes or breaks the performance of hardware-backed APIs like this is the amount of optimization behind the driver, and the ability of the application programmer to use the API in an efficient way. But coming up with two programs in two different APIs that provably do the exact same work using different APIs is almost impossible, because there are inevitably going to be mismatches in the APIs. There is not a clear one-to-one comparison between the two APIs, even though they purport to be able to accomplish the same kinds of tasks (i.e. general purpose parallel computing on a fully programmable GPU). You would also likely find that, for every instance of an app that runs like dog shit on one API, you can find an instance of another app that runs like dog shit on the other API. If you pay enough attention to optimization, algorithms, workarounds of slow routines, and so forth, you can pretty much make just about any task look really good on either API. It's a question of putting in the time to do that.

All that having been said, I wouldn't be surprised if CUDA is more heavily optimized in Nvidia's binary drivers than OpenCL. So you may be able to generally see a trend of faster results with CUDA, simply because Nvidia is more heavily invested in this technology, with it being of their own invention. This is also why Direct3D tends to perform better on Windows than OpenGL.

[dyna]

The GeForce 9800 GTX has more than twice the processing power and close to twice the bandwidth of the GeForce GT 220. If the GeForce 9800 GTX is losing to the GeForce GT 220, then either something is wrong or some sort of design innovation occurred that makes it faster. It would be nice if the review discussed this.

You mean something like this? As described on the last page...

Obviously these results are not hugely surprising with NVIDIA's newer generation of GPUs performing much better at GPGPU/OpenCL tasks than earlier generations.

[baryluk]

Huge numbers I see.

I'm not very surprised, as I quite often write in OpenCL and test it on Linux without problem, and it was fast. But how about actually comparing it to same benchmarks and hardware but using MS Windows? This will allow real comparission.

[bug77]

You mean something like this? As described on the last page...

Obviously these results are not hugely surprising with NVIDIA's newer generation of GPUs performing much better at GPGPU/OpenCL tasks than earlier generations.

The only leap in GPGPU performance since 8000 series was the GTX480/470/465. GTX460 had its GPGPU capability scaled back. And the 500 series apparently got the enhancements back. But when you're looking at 9000 vs 200 series, it's the same architecture. I think only the video engine was changes, besides the number of pipes.

[evolution]

Obviously these results are not hugely surprising with NVIDIA's newer generation of GPUs performing much better at GPGPU/OpenCL tasks than earlier generations. Even the low-end GeForce GT 220 graphics card is dramatically better with such workloads than a GeForce 9800GTX.

Really impressive these results. Latest Fermi-based cards put older nVidia cards on his dust! These new-gen Fermi cards really have a lot of GPGPU room...!

Latest "Fermi"-generation cards have a much higher number of dedicated GPGPU cores into their PCBs, so that they'll have very decent performance with OpenCL and other similar technologies (CUDA), OC!

Now, I'll be waiting some more time to know if AMD has a proper "answer" for these cards, but ATM, (and for some time) nVidia has been winning the GPGPU battle...

Cheers

[bug77]

Really impressive these results. Latest Fermi-based cards put older nVidia cards on his dust! These new-gen Fermi cards really have a lot of GPGPU room...!

Latest "Fermi"-generation cards have a much higher number of dedicated GPGPU cores into their PCBs, so that they'll have very decent performance with OpenCL and other similar technologies (CUDA), OC!

Now, I'll be waiting some more time to know if AMD has a proper "answer" for these cards, but ATM, (and for some time) nVidia has been winning the GPGPU battle...

Cheers

Do you even know what you're talking about? GT2xx is NOT Fermi.