GravityMark OpenGL/Vulkan Performance For NVIDIA RTX 30 vs. AMD Radeon RX 6000 Series

[reba]

This computer has in fact three Vulkan devices (one in software, two on hardware):

VK_LAYER_MESA_device_select (Linux device selection layer) Vulkan version 1.2.73, layer version 1:
Layer Extensions: count = 0
Devices: count = 3
GPU id = 0 (Intel(R) HD Graphics 4000 (IVB GT2))
Layer-Device Extensions: count = 0

GPU id = 1 (llvmpipe (LLVM 11.0.1, 256 bits))
Layer-Device Extensions: count = 0

GPU id = 2 (AMD RADV VERDE (ACO))
Layer-Device Extensions: count = 0

Letting it run with lavapipe gave me a segfault:

M: 0 us: GravityMark Started
M: 10.10 ms: ../data.zip: 313 files
M: 10.49 ms: Temporal antialiasing
WARNING: radv is not a conformant vulkan implementation, testing use only.
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
M: 1.053 s: Build Date: Jun 18 2021
M: 1.053 s: Build Info: release; vk=1; gl=45; gles=32
M: 1.090 s: System: Linux 5.12.10-xanmod1 x86_64 GNU/Linux
M: 1.090 s: Kernel: #0~git20210610.d262d74 SMP PREEMPT Thu Jun 10 13:59:27 UTC 2021
M: 1.090 s: Memory: 15.52 GB
M: 1.090 s: Uptime: 4 days 13:47
M: 1.090 s: CPU: Intel(R) Core(TM) i7-3635QM CPU @ 2.40GHz
M: 1.090 s: Device: VEN_8086&DEV_0166&SUBSYS_C0E6144D
M: 1.090 s: GPU: [AMD/ATI] Venus XT [Radeon HD 8870M / R9 M270X/M370X]
M: 1.090 s: Device: VEN_1002&DEV_6821&SUBSYS_C0E6144D
M: 1.090 s: Memory: 2.00 GB
M: 1.096 s: Desktop: 1920x1080 1.0
M: 1.096 s: Screen 0: 1920x1080 0 0 eDP-1
M: 1.096 s: Creating 1600x900 Vulkan Window
WARNING: lavapipe is not a conformant vulkan implementation, testing use only.
M: 1.110 s: Using Fetch mode
M: 2.326 s: Device: llvmpipe (LLVM 11.0.1, 256 bits)
M: 2.326 s: MaxViewportCount: 16
M: 2.326 s: MaxUniformSize: 64.00 KB
M: 2.326 s: MaxStorageSize: 128.00 MB
M: 2.326 s: Creating SceneManager
M: 6.579 s: Creating RenderManager
M: 7.926 s: Creating Scene
./run_windowed_vk.sh: line 30: 3121333 Segmentation fault ./GravityMark.x64 -temporal 1 -fps 1 -info 1 -sensors 1 -benchmark 1 -vk -device 1

It seems to crash in lavapipe:

Thread 38 "GravityMark.x64" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fff5defd700 (LWP 3128586)]
0x00007fffe391b0c1 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so

(gdb) where
#0 0x00007fffe391b0c1 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#1 0x00007fffe38b90b9 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#2 0x00007fffe38b972c in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#3 0x00007fffe38ba2c0 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_lvp.so
#4 0x00007ffff7a82878 in ?? () from ./libTellusim_x64.so
#5 0x00007ffff790e2f5 in ?? () from ./libTellusim_x64.so
#6 0x00007ffff790ed65 in ?? () from ./libTellusim_x64.so
#7 0x00007ffff77b9db2 in ?? () from ./libTellusim_x64.so
#8 0x00007ffff77c7176 in ?? () from ./libTellusim_x64.so
#9 0x00007ffff7245ea7 in start_thread () from /lib/x86_64-linux-gnu/libpthread.so.0
#10 0x00007ffff7153def in clone () from /lib/x86_64-linux-gnu/libc.so.6

Letting it run with the Intel device also gave me a segfault:

M: 0 us: GravityMark Started
M: 11.97 ms: ../data.zip: 313 files
M: 12.50 ms: Temporal antialiasing
WARNING: radv is not a conformant vulkan implementation, testing use only.
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete
M: 1.066 s: Build Date: Jun 18 2021
M: 1.066 s: Build Info: release; vk=1; gl=45; gles=32
M: 1.089 s: System: Linux 5.12.10-xanmod1 x86_64 GNU/Linux
M: 1.089 s: Kernel: #0~git20210610.d262d74 SMP PREEMPT Thu Jun 10 13:59:27 UTC 2021
M: 1.089 s: Memory: 15.52 GB
M: 1.089 s: Uptime: 4 days 14:15
M: 1.089 s: CPU: Intel(R) Core(TM) i7-3635QM CPU @ 2.40GHz
M: 1.089 s: Device: VEN_8086&DEV_0166&SUBSYS_C0E6144D
M: 1.089 s: GPU: [AMD/ATI] Venus XT [Radeon HD 8870M / R9 M270X/M370X]
M: 1.089 s: Device: VEN_1002&DEV_6821&SUBSYS_C0E6144D
M: 1.089 s: Memory: 2.00 GB
M: 1.094 s: Desktop: 1920x1080 1.0
M: 1.094 s: Screen 0: 1920x1080 0 0 eDP-1
M: 1.094 s: Creating 1600x900 Vulkan Window
M: 1.145 s: Using Fetch mode
M: 1.362 s: Device: Intel(R) HD Graphics 4000 (IVB GT2)
M: 1.362 s: MaxViewportCount: 16
M: 1.362 s: MaxUniformSize: 128.00 MB
M: 1.362 s: MaxStorageSize: 1.00 GB
M: 1.362 s: Creating SceneManager
Segmentation fault

And it seems to crash in Intel's Vulkan driver, but on a different path of Tellusim:

Thread 1 "GravityMark.x64" received signal SIGSEGV, Segmentation fault.
0x00007fffeb0cbfe5 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so

(gdb) where
#0 0x00007fffeb0cbfe5 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#1 0x00007fffeb0ce8d5 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#2 0x00007fffeb037245 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#3 0x00007fffeb02d748 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#4 0x00007fffeaf41e36 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#5 0x00007fffeaf5d0e4 in ?? () from /usr/lib/x86_64-linux-gnu/libvulkan_intel.so
#6 0x00007ffff7a804e8 in ?? () from ./libTellusim_x64.so
#7 0x00007ffff792c7f7 in ?? () from ./libTellusim_x64.so
#8 0x00007ffff79212fb in ?? () from ./libTellusim_x64.so
#9 0x00007ffff7920bac in ?? () from ./libTellusim_x64.so
#10 0x00007ffff7782416 in Tellusim::SceneManager::create(Tellusim::Device const&, Tellusim::SceneManager::Flags, Tellusim::Function<void (unsigned int)> const*) () from ./libTellusim_x64.so
#11 0x000000000040d95b in ?? ()
#12 0x000000000041034e in ?? ()
#13 0x000000000040b31d in ?? ()
#14 0x00007ffff707cd0a in __libc_start_main () from /lib/x86_64-linux-gnu/libc.so.6
#15 0x000000000040b15a in ?? ()

Any ideas?

[Termy]

Small typo:

AMD RDNA2 performance was coming in short of expectations with the Radeon RX 6800 XT only hitting the NVIDIA RTX 3080 Ti.

should probably be 3070 TI

Interesting to see such a stark contrast between vulkan and openGL in regards to the vendors - makes me question the usefulness of the benchmark as a whole when the main selling point is comparing vulkan/openGL ^^

[smitty3268]

This computer has in fact three Vulkan devices (one in software, two on hardware):

GPU id = 0 (Intel(R) HD Graphics 4000 (IVB GT2))

WARNING: lavapipe is not a conformant vulkan implementation, testing use only.
MESA-INTEL: warning: Ivy Bridge Vulkan support is incomplete

Any ideas?

Not to be too pessimistic, but lavapipe and Ivy Bridge vulkan drivers are both known to be incomplete and buggy. Not a huge surprise they might run into issues.

[middy]

Small typo:

should probably be 3070 TI

Interesting to see such a stark contrast between vulkan and openGL in regards to the vendors - makes me question the usefulness of the benchmark as a whole when the main selling point is comparing vulkan/openGL ^^

i personally think its coming down to optimizations. since this is the first release for linux, i'm sure there is no mesa optimizations for it at all. while with nvidia, there probably has been due to it already being on windows. i would be curious to see how the close source amd drivers perform but i don't think it would help much, as i remember reading here awhile ago from a AMD rep that the linux drivers, close and open (especially open as mesa is used), don't share much with their windows driver. i could be wrong though.

[reba]

Not to be too pessimistic, but lavapipe and Ivy Bridge vulkan drivers are both known to be incomplete and buggy. Not a huge surprise they might run into issues.

True. But maybe GravityMark is using something, which makes the two fall over, that's kind of optional and thus enables lavapipe and intel by removing it.

[Yoshi]

here are my 1080p results with my 6900xt for those curious:

vulkan: https://i.imgur.com/rn86Hwg.png
opengl: https://i.imgur.com/93pZrqM.png

specs:
amd ryzen 5800x
32gb ddr4 3200 cas 16
amd reference 6900xt
msi b550 tomahawk with latest agesa
sam enabled
arch linux - kernel 5.12.11-arch1-1
mesa 21.1.2-1
radv 21.1.2-1

vulkan: https://abload.de/image.php?img=grav...k_0036pkoz.png
opengl: https://abload.de/image.php?img=grav...k_004zmj3g.png

specs:
amd ryzen 5950x (water cooled)
64gb ddr4 4000 cas 19
amd reference 6900xt (water cooled)
gigabyte x570 aorus master f33 (water cooled)
sam enabled
windows 10 19043.1055
amd driver 21.5.2

[frustum]

Small typo:

should probably be 3070 TI

Interesting to see such a stark contrast between vulkan and openGL in regards to the vendors - makes me question the usefulness of the benchmark as a whole when the main selling point is comparing vulkan/openGL ^^

The main purpose of the benchmark is to showcase the driver problems. Otherwise, everything is fine for the driver vendors.

It should be no difference between Vulkan and OpenGL because the benchmark is not using multi-threading, and a single API call performs all rendering.
But nobody about OpenGL anymore. The difference between Vulkan running Linux and Windows should be zero. But there is a difference, and Linux is not winning.

Intel Mesa driver only works under OpenGLES where multi_draw_indirect_count() is emulated with a loop of multiple draw_indirect().
OpenGLES provides 20+ FPS. Vulkan and OpenGL are not capable of going over 1 FPS.

One of the main purposes of the Vulkan was multiple-GPU support. It's not available under Linux with Nvidia. That why the next benchmark feature will be multi-GPU support under Windows. And the performance of Linux will be just busted.

[Linuxxx]

The main purpose of the benchmark is to showcase the driver problems. Otherwise, everything is fine for the driver vendors.

It should be no difference between Vulkan and OpenGL because the benchmark is no using multi-threading, and a single API call performs all rendering.
But nobody about OpenGL anymore. The difference between Vulkan running Linux and Windows running should be zero. But there is a difference, and Linux is not winning.

Intel Mesa driver only works under OpenGLES where multi_draw_indirect_count() is emulated with a loop of multiple draw_indirect().
OpenGLES provides 20+ FPS. Vulkan and OpenGL are not capable of going over 1 FPS.

One of the main purposes of the Vulkan was multiple-GPU support. It's not available under Linux with Nvidia. That why the next benchmark feature will be multi-GPU support under Windows. And the performance of Linux will be just busted.

Could You please share which Linux distro You used for the comparison against Windows, and whether You were comparing both in their out-of-the-box states?

[frustum]

Could You please share which Linux distro You used for the comparison against Windows, and whether You were comparing both in their out-of-the-box states?

This is a comparison between Windows and Linux on GeForce 1660:
https://www.techpowerup.com/forums/t...2#post-4543860

This is the difference (15%) between xubuntu 18.04 (without composition) and Windows 10 running the same HW:

https://tellusim.com/upload/rtx2080ti_linux_windows.jpg

[theriddick]

I guess this just shows that Vulkan AMD drivers still need allot of work under Linux. I don't know how things look under windows, possibly they struggle there also.