Fresh AMDGPU-PRO vs. NVIDIA OpenCL LInux Benchmarks - 11 GPUs

[phoronix]

Phoronix: Fresh AMDGPU-PRO vs. NVIDIA OpenCL LInux Benchmarks - 11 GPUs

While running the fresh NVIDIA vs. AMD Vulkan Linux benchmarks (that also included some OpenGL numbers too), I had also taken the opportunity to run some fresh OpenCL compute benchmarks of the latest NVIDIA 370 proprietary Linux driver against AMDGPU-PRO on different graphics cards...

http://www.phoronix.com/scan.php?page=news_item&px=October-2016-11-GPU-CL

[adakite]

Are these tests FP32 only? Could it be possible to mention the respective precision used for each bench? Thanks

[Jumbotron]

I"ll repeat my request here from the Nvidia vs AMD showdown. I'd like to see a showdown between different generations of AMD APUs. Mullins, Jaguar, Kaveri, Carrizo and even the new Bristol Ridge.

Also.....what effects different speed DDR RAM has on performance. Like the difference between DDR 3 1600 vs DDR 3 2400 or even DDR 4 2400 since Bristol Ridge can use DDR 4 like the upcoming ZEN APU.

[pcxmac]

Would be nice to see some windows benchmarks too, see how bad the drop is.

[mannerov]

It's very easy to favor a vendor over another:
. use fp64 on one, and fp32 (or even fp16) on the other (huge perf impact)
. use optimized shader for one vendor, and generic shader on another (optimizing an opencl shader for an architecture can be x2 or more gains)

Unfortunately, it just seems to show some of these benchmarks had some nvidia optimizations.

[ElectricPrism]

It's very easy to favor a vendor over another:
. use fp64 on one, and fp32 (or even fp16) on the other (huge perf impact)
. use optimized shader for one vendor, and generic shader on another (optimizing an opencl shader for an architecture can be x2 or more gains)

Unfortunately, it just seems to show some of these benchmarks had some nvidia optimizations.

Or fortunately depending on how you looka at it. I've always speculated that benchmarks can at the very least tip the odds towards at least some in one camp's favor although I'm sure they put some effort into being neutral.

So what is the ELI5 message of these benchmarks? And the driver tested has the same OpenCL version support and so on as the counterparts in the driver?

[patstew]

It's very easy to favor a vendor over another:
. use fp64 on one, and fp32 (or even fp16) on the other (huge perf impact)
. use optimized shader for one vendor, and generic shader on another (optimizing an opencl shader for an architecture can be x2 or more gains)

Unfortunately, it just seems to show some of these benchmarks had some nvidia optimizations.

Using a different FP precision would be massively dishonest. 'Optimisation' is much harder to pin down, to what extent is it tweaked an to what extent is it just a workload that favours the strengths of a particular chip?. Different algorithms stress different things, so you need to pick one that's representative of the workload you're interested in.

[oooverclocker]

The only OpenCL test cases where AMD came out far ahead was in LuxMark, which tends to favor AMD hardware/drivers.

Fury X: 8.6 TFLOPS SP
GTX 980 Ti: 6.1 TFLOPS
RX 480: 5.8 TFLOPS

Where do you see Luxmark favoring anyone? The Luxmark results are exactly as expected. With the Exception of the GTX 1070 and 1080.
The SHOC benchmark results show mostly unexpected behavior.

[smitty3268]

Fury X: 8.6 TFLOPS SP
GTX 980 Ti: 6.1 TFLOPS
RX 480: 5.8 TFLOPS

Where do you see Luxmark favoring anyone? The Luxmark results are exactly as expected. With the Exception of the GTX 1070 and 1080.
The SHOC benchmark results show mostly unexpected behavior.

Just a reminder, Michael is testing a Fury (non-X) not the Fury X.

[Amarildo]

Would be nice to see some windows benchmarks too, see how bad the drop is.

As far as the real world goes, I've seen no difference.