NVIDIA GeForce GTX 1080 On Linux: OpenGL, OpenCL, Vulkan Performance

[rabcor]

Very impressive, can't wait to see the 1070 and the radeon 400 series added to the lineup for a comparison.

[rhysk]

Thanks for the review Michael.

[Zan Lynx]

Yes thanks!

[Michael]

Michael
Thanks for the review

But one question, did you do the benchmarks in an open or closed rig?
I'm aware that GTX 1080 runs into throttling quicker than it's predecessors, and usually quickly falls down to it's base frequency within 2-4 minutes of load. So to get reproducible results the card should have a "warm up" before the benchmark itself, or run multiple times.

Open. But I guess you should read more Phoronix :P The Phoronix Test Suite already runs every test a minimum of three times. If the standard deviation is above 3.5%, PTS automatically re-runs the test more times (up to 2x times however many times the test profile told it to run to begin with, so usually six times). Any deviation is also shown on the result graphs themselves.

[Nille_kungen]

I liked the clinfo, vdpauinfo and glxinfo and the screenshot shows the pci-id, that is something i like to see on new hardware.
About the open driver on AMD i rather see the open drivers since that's what i use but some might say that it would be more fair test if tested with amd blob.

[Michael]

I liked the clinfo, vdpauinfo and glxinfo and the screenshot shows the pci-id, that is something i like to see on new hardware.
About the open driver on AMD i rather see the open drivers since that's what i use but some might say that it would be more fair test if tested with amd blob.

FYI, pretty much for all of my tests I upload the system logs with my benchmarks to OpenBenchmarking.org. So it's really nothing new, but just that I featured it prominently this time, as generally speaking for most benchmark results on OpenBenchmarking.org with a few mouse clicks you can pull up all of the information.

[bridgman]

No real reason besides when I did the recent beta 2 vs. Linux 4.6 / Mesa 11.3-dev tests the OpenGL results were rather close...

Makes sense... it's really Vulkan and OpenCL I was thinking about... certainly Vulkan does a better job than OpenGL of showing what the relative performance is going to be like on upcoming games.

And then in some past comparisons when I've tested with Catalyst vs. NVIDIA, people complained I should have used the open-source driver instead. So now I decided to use open-source.

Yep, agreed... AFAICS what people really want is open source OpenGL plus the hybrid stack's Vulkan and OpenCL.

But I guess people will complain either way.

Sometimes it's even the same people

If there's enough interest, happy to run a GTX 1080 comparison with AMDGPU-PRO, or will certainly use both drivers when it comes to RX 480 testing.

Great, thanks. This might actually work out well since the next amdgpu hybrid driver should be coming out fairly soon.

[Michael]

Makes sense... it's really Vulkan and OpenCL I was thinking about... certainly Vulkan does a better job than OpenGL of showing what the relative performance is going to be like on upcoming games.

Great, thanks. This might actually work out well since the next amdgpu hybrid driver should be coming out fairly soon.

Yeah had I been running more OpenCL and Vulkan benchmarks, definitely would have used -PRO then for this article, but mostly was focusing on OpenGL and then just Dota 2 for Vulkan.

Great to hear a new hybrid will be coming out soon.... Can't wait to run some fresh tests!

[artivision]

Any compute device that has preemptive multithreading has asynchronous capabilities by definition. In the Pascal generation, NVidia has introduced much finer-grained preemption, so it's ability to handle asynchronous instructions has improved substantially. Is the hardware support as effective as AMD's implementation with their ACE's? Well, that's another argument for another day, and I don't know the answer. My guess in some cases it will be better, and other cases, like an-AMD-funded-and-biased-benchmark-posing-as-real-game-aka-AOTS, it will not. So it's fair to say the AMD have used a different means to support asynchronous compute, and even to argue that it's better. But to say NV doesn't "have" it is incorrect.

Also, remember that AMD may have paid a huge price for those ACE engines. Notice how the latest NV hardware gets 1.8GHz on air, while the Vega is pegged at 1.2GHz. It's like CISC vs. RISC. Yes, AMD may have better hardware implementation of complex logic, but all that overhead may require sacrifices in clock speed, ROPs[1], etc. that in the end result in poorer performance for most use cases. Look at the apps you want to run on your OS and then the performance of your card for those apps, *then* buy your card. If I were to use simple-minded metrics like "it ain't got no async_compute like AMD" I would have a Fury X which provides less than half the performance of my GTX980Ti at the same price and many times the number of headaches. Yeah, no thanks.

EDIT:
[1] It appears Vega has only 64 ROPs, same as the Fury X. The GM204 products have 96 ROPs; the GP102 products (1080Ti, Titan P?) will probably have a similar number.

Completely nothing to do with efficiency. Async_compute means to do compute upon and during shading (aka complex shaders). Standard compute means to do compute after shading on the entire shader. This is not a hardware nor an Api future, it's developers thinking. Nvidia hardware runs this but doesn't accelerate it (the gain in beauty of the field compensates for the loss anyway). With Amd there is no loss at all. Don't find it strange if a future FX (a modern Motion Blur for example) cannot even run on an Nvidia Gpu at all and give something like 40ms latency. Amd has this magnificent blend threading but also good connection-compression of information. In Ashes TitanX is just as fast as a 390x and TitanX cheats with snow and other FX.

[eydee]

*bridgman scratches head, tries to understand why a comparison with a new high end card requires the inclusion of older AMD cards to be relevant, or why every card has to run the same driver. Not succeeding yet.

I'm sure Michael would be very happy if AMD sent out free samples of more recent cards to benchmark. He can use what he has, can't keep up with all the rebranding and batch moving to legacy... Feel free to contribute.