Radeon Software 18.40 vs. Mesa vs. AMDVLK Benchmarks With Radeon RX Vega

[phoronix]

Phoronix: Radeon Software 18.40 vs. Mesa vs. AMDVLK Benchmarks With Radeon RX Vega

This week marked the release of Radeon Software 18.40 as the latest release of AMD's Linux driver stack targeting workstation users. While the sole mentioned change was the addition of SUSE Linux Enterprise 15 support, I decided to run some benchmarks of this latest driver compared to the other open-source Radeon Linux driver options.

http://www.phoronix.com/vr.php?view=27027

[darkbasic]

Incredible, AMDVLK is starting to beat RADV...

[TemplarGR]

1. Civilization VI does not officially support non-Nvidia blob drivers for the Linux port. As you can see from the benchmark, the performance is utterly abysmal. Attrocious framerates even at the lowest settings, on the powerful Vega 56 gpu... On my R9 380 it is even worse, it is literally unplayable at lowest settings @ 1050p, while on Windows i max out all the settings and it runs fine. Most of the time it fails to even load, and crashes X for me. Garbage port.

2. As we can see in Rise of the Tomb Raider, RADV has a lot of room for performance optimizations. I appreciate the work done on the driver but seeing how much better AMDVLK has become perhaps it is time to just drop it like Radeon-HD and embrace AMDVLK. The ideal would be for current RADV developers to work together with AMDVLK developers instead of duplicating effort.

[marek]

1. Civilization VI does not officially support non-Nvidia blob drivers for the Linux port. As you can see from the benchmark, the performance is utterly abysmal. Attrocious framerates even at the lowest settings, on the powerful Vega 56 gpu... On my R9 380 it is even worse, it is literally unplayable at lowest settings @ 1050p, while on Windows i max out all the settings and it runs fine. Most of the time it fails to even load, and crashes X for me. Garbage port.

Last time I checked, Civ VI had the same performance on RadeonSI as on NVIDIA.

[jrch2k8]

1. Civilization VI does not officially support non-Nvidia blob drivers for the Linux port. As you can see from the benchmark, the performance is utterly abysmal. Attrocious framerates even at the lowest settings, on the powerful Vega 56 gpu... On my R9 380 it is even worse, it is literally unplayable at lowest settings @ 1050p, while on Windows i max out all the settings and it runs fine. Most of the time it fails to even load, and crashes X for me. Garbage port.

2. As we can see in Rise of the Tomb Raider, RADV has a lot of room for performance optimizations. I appreciate the work done on the driver but seeing how much better AMDVLK has become perhaps it is time to just drop it like Radeon-HD and embrace AMDVLK. The ideal would be for current RADV developers to work together with AMDVLK developers instead of duplicating effort.

Andser for 2.)

1.) AMDVLK still fails a freaking lot, again michael handpick those games prolly cuz it runs on both, otherwise you cannot run a comparison.

2.) RADV is leaps and bounds better on wayland

3.) AMDVLK is barely usable on DXVK and VKD3D at least on PRE-VEGA cards(even my TAHITI 7950 refresh runs Wolfenstein and DOOM at ultra 1080p butter smooth with RADV + a serious bunch of DX11 titles)

4.) RADV integrates fully with mesa and share a lot of code with Intel ANV driver(hence a lot more eyes fixing bugs)

5.) RADV and OpenGL(4.6) integration is around the corner, not sure will be so easy on AMDVLK

6.) There is no duplication of efforts, AMDVLK is basically a semi regular Linux compatible code dump from Windows Driver while RADV is a full community effort and is totally integrated into Mesa, if you are sick of 2 driver for some reason(there is no none BTW, RADV is for Gamers while AMDVLK will end up for either pro users or LTS distros like Red Hat where is not easy to upgrade Mesa/LLVM) then forget about AMDVLK since overall is the inferior driver(outside of FPS junkies that jump every time some game show 10fps gain and completely disregard everything else)

7.) RADV is the defacto driver for CodeWeavers and Valve and honestly the difference between drivers is so small that i don't think they will even consider change their position about it, specially since few pending fixes on LLVM and NIR backends will improve this fairly soon

8.) OpenCL over NIR/HMM seems feasible enough at this point and this could bring even more benefits for RADV/RadeonSI in the mid term that doesn't look that easy for AMDVLK to reach at this point since their OpenCL code is a mess to say the least right now depending on your GPU.(DAL vs ROCM vs GCN versions)

[nuetzel]

With the aging Unigine Heaven OpenGL performance benchmark, AMDGPU-PRO 18.40 surprisingly is still much faster than RadeonSI Gallium3D.

'Old' AMDGPU-PRO 'hand optimization' (special path)?!

[Marc Driftmeyer]

Run the Blender Benchmarks. Let's compare performance of the OpenCL stack on the GPGPU drivers. Oh wait!

[humbug]

Great performance by the proprietary Vulkan driver + compiler !!
Beating RadV by 20% in Rise of the Tomb Raider and 30% in Talos Principle. These are big margins.

Vega is currently the fastest AMD graphics part so it can serve to highlight the differences in efficiency of the software stack.

If the open stack had that kind of vulkan performance it would make vega a more appealing buy for Linux gamers.

[TemplarGR]

Last time I checked, Civ VI had the same performance on RadeonSI as on NVIDIA.

You should check again then. Do you see Michael' benchmarks? Do you really think that Civ VI should be getting this low framerates on top hardware? Civ VI is not exactly a demanding game graphically...

[Big_Mama]

"Gallium AMD-GPU PRO SI Open Mesa Radeon 3D?"

"No, no!... Nvidia!"