NVIDIA GeForce GTX 980 Dominates With OpenCL On Linux

[Marc Driftmeyer]

Why in the hell are you compiling OpenCL tests with OpenMP enabled?

[Marc Driftmeyer]

What is the highest OpenCL version NVidia currently supports?

Don't tell me it's still OpenCL 1.1 only.

Yes and this test suite sure as hell isn't OpenCL 1.2 never mind OpenCL 2.0. But compiling these results with OpenMP enabled is asinine.

The non-OpenMP test is land locked at OpenCL 1.1.

http://openbenchmarking.org/innhold/066ebb00270596313732d697ca0091d9649546b2

Move these tests to OpenCL 1.2 compliance and you won't see an Nvidia card in the list.

[Sdar]

Yes and this test suite sure as hell isn't OpenCL 1.2 never mind OpenCL 2.0. But compiling these results with OpenMP enabled is asinine.

The non-OpenMP test is land locked at OpenCL 1.1.

http://openbenchmarking.org/innhold/066ebb00270596313732d697ca0091d9649546b2

Move these tests to OpenCL 1.2 compliance and you won't see an Nvidia card in the list.

Move these tests to large/complex opencl kernels and you won't see an AMD card in the list.

[guido12]

If you're only concerned about OpenCL/CUDA and not gaming, what are the downsides of getting a GeForce card instead of a Quadro? What if I mainly need 64-bit floating point performance? Or what if 32-bit floats are good enough?

[Kivada]

I'm actually probably going to wait for the Pascal line for my next "big" card. I have a 770 which still does just about everything I throw at it for Linux. Even on Windows, it still isn't pushed around by pretty much anything.
I'm thinking we're coming up on a period of time where graphics aren't going to be able to keep up with hardware instead of the other way around for once. I look forward to a time where I can have cool looking graphics with something that isn't running 100% all the time that heats up my entire room.

Every time we hit that point a new generation of game consoles come out, the new consoles are now out and they are using 8 core CPUs with OpenGL4.5 class GPUs, instead of the equivalent of 1-3 cores with OpenGL2.1 class GPUs. This allows game dev to spread their wings again by raising the bare minimum bar they have to target.

[Michael]

Why in the hell are you compiling OpenCL tests with OpenMP enabled?

You're just talking about Parboil, right? It shows up that way in Parboil since Parboil has several different versions of each test -- when run on the CPU it uses OpenMP, when run on the GPU there's OpenCL, and I think there's also a CUDA version. Due to the several different binaries produced, PTS just sees the first one when it's monitoring the compiler flags being tested -- so it reports for Parboil as the OpenMP flag since the CPU version (the most common Parboil test case) is compiled for the CPU. But for this article, only the OpenCL version is tested.

[justmy2cents]

I don't get your argument.
This sounds like you're blaming AMD for NVidia not having OpenCL >= 1.2 support.

We all know it's just a political thing NVidia does not provide support for newer OpenCL version.
And this is bad. No matter how great their OpenCL 1.1 support is.


Wait, I thought many problems arise because AMDs OpenGL implementation is stricter wrt the standard.

i wouldn't say he blamed anything. amd opencl is broken and that is a known fact you can check on blender for example

[mitcoes]

Thanks, but I always miss a price/performance chart

Lets say more or less the GTX 750 Ti is half as good as the 780 Ti and a third of good as the 980

and actual prices are 160 USD 566 USD and unknown so 2 GTX 750 Ti in SLI that would cost near half a 780 Ti perhaps would almost tie with it (or not) and It is worth to benchmark just because of price/performance benchmark that we usually miss at phoronix reviews.

[curaga]

Wait, I thought many problems arise because AMDs OpenGL implementation is stricter wrt the standard.

I can only speak for GL, but he's right. The strict-GL-compliance is one set of issues, but GL functionality breaking on the AMD blob is also a big issue. For fucks sake, half the time even standard GL2 functionality is broken on the blob, while many of the new 4.4 things take months to start working *after* AMD announces they're fully supported.

Nowadays I only target mesa on AMD cards. Not only is it even stricter than the AMD blob, it doesn't have terrible surprising regressions in version X + 1. NV blob is perfectly fine in these regards, regressions rarely, and when NV claims new functionality, it works 95% of the time.

[tuke81]

Lets say more or less the GTX 750 Ti is half as good as the 780 Ti and a third of good as the 980

and actual prices are 160 USD 566 USD and unknown so 2 GTX 750 Ti in SLI that would cost near half a 780 Ti perhaps would almost tie with it (or not) and It is worth to benchmark just because of price/performance benchmark that we usually miss at phoronix reviews.

gtx750ti does not support sli(a bit moot point in gpgpu because they don't need to be slied anyway).

I agree with you though, price/performance chart would be useful. GTX 980 is not the card to get, the gtx970 is(price/performance point of view).