1080p/1440p Linux Gaming Performance For Radeon RX 590/Vega & NVIDIA 1060/1070/1080/2060/2070

[deppman]

Michael
Curious why you did this: "On the NVIDIA driver side was the 395.25 driver atop the Linux 4.20 kernel". The latest driver is Nvidia 415.25. My guess is this wasn't compatible with the kernel, but shouldn't it be used instead? I expect the NV results may have been a bit better especially on the RTX cards with the latest driver.

[Michael]

@Michael

As other said many thanks for 1080p tests

However can add in future grid autosports, shadow of mordor and batman (dxvk)

Another question is possible test outlast for verify amd cards stay affected in unreal engine 3 games ?

At least is possible show shadow of mordor benchmark 1080p using same card but using various cpus: intel - amd for see how ghz depend in this game ?

Thanks

GRID Autosports and Shadow of Mordor have no way to automate benchmarks from the CLI hence why they are not tested besides my manual tests of them once every few years.

[pinguinpc]

@Michael

As other said many thanks for 1080p tests

However can add in future grid autosports, shadow of mordor and batman (dxvk)

Another question is possible test outlast for verify amd cards stay affected in unreal engine 3 games ?

At least is possible show shadow of mordor benchmark 1080p using same card but using various cpus: intel - amd for see how ghz depend in this game ?

Thanks

[Strunkenbold]

I don't think this is a regression. I've always had this problem on Ultra preset.
https://bugs.freedesktop.org/show_bug.cgi?id=100069

Yup. This bug has never been fixed. But since Michael is always testing with different settings, the problem didn't showed up.
I have up the hope that one of the Radeonsi guysg fix the problem. Even though I think it's worth fixing the bug. If appears to be a general problem with the ego engine as Dirt3 via dxvk is also painfully slow at Max settings.

[Twysock]

Am bewildered at how some of the GTX and RTX bench graphs are nearly neck and neck @1080p. For sure Vega 56 is a good competitor for the GTX 1070 which is my card but 56 has a notably higher power draw. I wonder if Vega will get a refresh with lower TDP.

Per the rumor mill, it won't have one. Just Vega VII, and then Navi sometime in the second half. Where Navi will perform/price is the big question.

[haagch]

It's done when I switch to a non-GNOME desktop for testing since otherwise CSGO is still sadly stuck with GNOME thinking the application is unresponsive unless I click it right away, otherwise it would affect results.

CSGO is unresponsive. When starting the game it goes unresponsive for at least 40 seconds. It also compiles about 3000 shaders in that time, but doesn't use a lot of CPU.
It also is unresponsive when loading any map for at least 14 seconds without any high CPU or hard disk usage.
Since recently csgo also randomly segfaults.

The windows version of CSGO in wine does not have these issues. The game itself and maps load much faster and it doesn't crash. But on Vega with nine it causes random kernel bugs:

 general protection fault: 0000 [#1] PREEMPT SMP NOPTI
 CPU: 9 PID: 18604 Comm: csgo.exe Not tainted 5.0.0-rc1-mainline #1
 Hardware name: System manufacturer System Product Name/PRIME B350-PLUS, BIOS 4023 08/20/2018
 RIP: 0010:__memcpy+0x12/0x20
 Code: ff 0f 31 48 c1 e2 20 48 09 c2 48 31 d3 e9 79 ff ff ff 90 90 90 90 90 90 0f 1f 44 00 00 48 89 f8 48 89 d1 48 c1 e9 03 83 e2 07 <f3> 48 a5 89 d1 f3 a4 c3 66 0f 1f 44 00 00 48 89 f8 48 89 d1 f3 a4
 RSP: 0018:ffffb76a60e17c70 EFLAGS: 00010246
 RAX: 9fcd16fedc2296ce RBX: 00000000d3e00000 RCX: 0000000000040000
 RDX: 0000000000000000 RSI: ffffa21b53e00000 RDI: 9fcd16fedc2296ce
 RBP: ffffa2227a6920b0 R08: ffffa21a80000000 R09: 0000000000000200
 R10: fffff74fd668cdc0 R11: ffffa21e36aa87a8 R12: 0000000000200000
 R13: 0000000000000000 R14: 0000000000000096 R15: ffffffffc081f7b8
 FS:  000000007ffd8000(0063) GS:ffffa2227ec40000(006b) knlGS:00000000f7ab1700
 CS:  0010 DS: 002b ES: 002b CR0: 0000000080050033
 CR2: 00000000abb86e00 CR3: 000000059634a000 CR4: 00000000003406e0
 Call Trace:
  dma_direct_unmap_page+0x7a/0x80
  ttm_unmap_and_unpopulate_pages+0xdb/0xf0 [ttm]
  ttm_tt_destroy.part.6+0x49/0x50 [ttm]
  ttm_bo_cleanup_memtype_use+0x2e/0x70 [ttm]
  ttm_bo_put+0x22a/0x280 [ttm]
  amdgpu_bo_unref+0x1a/0x30 [amdgpu]
  amdgpu_gem_object_free+0x33/0x50 [amdgpu]
  drm_gem_object_release_handle+0x59/0xc0 [drm]
  drm_gem_handle_delete+0x5e/0x90 [drm]
  ? drm_gem_handle_create+0x40/0x40 [drm]
  drm_ioctl_kernel+0xaf/0xf0 [drm]
  drm_ioctl+0x25f/0x3f0 [drm]
  ? drm_gem_handle_create+0x40/0x40 [drm]
  ? ktime_get+0x36/0xa0
  __ia32_compat_sys_ioctl+0x2f7/0xd10
  do_fast_syscall_32+0xa7/0x2a0
  entry_SYSCALL_compat_after_hwframe+0x45/0x4d

With these issues and its numerous random stutters both native and with nine I can't imagine that anyone would play csgo competitively on Linux.

We can only hope for an eventual Source 2 port that runs with Vulkan.

[creative]

Am bewildered at how some of the GTX and RTX bench graphs are nearly neck and neck @1080p. For sure Vega 56 is a good competitor for the GTX 1070 which is my card but 56 has a notably higher power draw. I wonder if Vega will get a refresh with lower TDP.

[Pentarctagon]

It looks like most of these aren't entirely GPU bottlenecked, or if they are that Nvidia is generally doing something that's a bit more performant.

[humbug]

The newest amdvlk release had general performance optimizations which should benefit most Vulkan applications. Would be interesting to see how it stacks up against radV now.

[Michael]

Thanks for including Counter-Strike: Global Offensive in your tests. Even though they were quite CPU bound, the benchmarks were revealing & useful.

It's done when I switch to a non-GNOME desktop for testing since otherwise CSGO is still sadly stuck with GNOME thinking the application is unresponsive unless I click it right away, otherwise it would affect results.