AMDGPU-PRO 17.20 Benchmarking vs. RadeonSI/RADV

[phoronix]

Phoronix: AMDGPU-PRO 17.20 Benchmarking vs. RadeonSI/RADV

Thanks to this week's Radeon Vega Frontier Edition launch, AMD pushed out a new build of their hybrid driver stack for Linux, AMDGPU-PRO. This new release is marketed as AMDGPU-PRO 17.20 and is only found when looking for the Frontier driver, but it's been working out fine so far in my Polaris/Fiji GPU testing. Here are some benchmarks compared to their current stable series, AMDGPU-PRO 17.10, as well as the newest open-source AMDGPU+RadeonSI/RADV driver stack.

http://www.phoronix.com/vr.php?view=24878

[tildearrow]

Thank you for Triangle, but it wasn't at 1920x1080... Typo:

its support seems to have regressed in the past week. WIth RadeonSI, the game was crashing during the loading screen.

[leipero]

At this point AMD should just add more developers contributing to Mesa and go with free drivers, I seriously doubt there could be any advantage in hiding something from competition with proprietary drivers, could be wrong.

[M@yeulC]

Interesting; RADV seems to score its first win against the PRO driver with SS3/Fusion. I am actually slightly worried for AMD and the future of its driver here...

The DX:MD regression is also a bit worrying, but hopefully only temporary.

Except than this, not much to be seen here. Great job, RadeonSI devs! It makes me want to buy a GCN card sooner than planned to enjoy the goodies as well

[agd5f]

At this point AMD should just add more developers contributing to Mesa and go with free drivers, I seriously doubt there could be any advantage in hiding something from competition with proprietary drivers, could be wrong.

It's not about hiding anything, it's about supporting workstation customers. For gaming we recommend mesa OpenGL.

[agd5f]

Note that the 17.20 release is specifically for Vega10 Frontier.

[Marc Driftmeyer]

I have no idea why you are testing Blender 2.78c when all the OpenCL performance gains are in the upcoming 2.79 nightlies. They are rock steady. Go grab it and re-test it.

[jrch2k8]

the good thing I see in these kind of benchmarks is the fact everyday RADV gets better and better, prolly by august - november AMD Vulkan driver would be an afterthought or for workstations only

[Michael]

I have no idea why you are testing Blender 2.78c when all the OpenCL performance gains are in the upcoming 2.79 nightlies. They are rock steady. Go grab it and re-test it.

Tried testing with 2.79 but they seemed to have changed their Python API as the benchmarking no longer works and at last check couldn't figure out a workaround/correction.

[microcode]

Phoronix: AMDGPU-PRO 17.20 Benchmarking vs. RadeonSI/RADV

Thanks to this week's Radeon Vega Frontier Edition launch, AMD pushed out a new build of their hybrid driver stack for Linux, AMDGPU-PRO. This new release is marketed as AMDGPU-PRO 17.20 and is only found when looking for the Frontier driver, but it's been working out fine so far in my Polaris/Fiji GPU testing. Here are some benchmarks compared to their current stable series, AMDGPU-PRO 17.10, as well as the newest open-source AMDGPU+RadeonSI/RADV driver stack.

http://www.phoronix.com/vr.php?view=24878

Cool benchmarks, but please try to get 99.9th percentile frame time (often represented as "0.1% low" by 1000 / (99.9th percentile frametime)). RadeonSI seems to have much better minimum framerates and max frametimes, which I think is far more important than comparing average framerates (especially when they're close on that metric).

A system with an average framerate of 120fps but a 99.9th percentile max frame time of 20ms is much worse than a system with 70fps average framerate, but 99.9th percentile max frame time of 16.4ms