Radeon RX 6800 Series Has Excellent ROCm-Based OpenCL Performance On Linux

[baryluk]

Michael, you can also add gpuowl and mandelbulber2 to the OpenCL tests. Both can be run from command line too. Maybe they are even in pts already. I don't know.

I.e. I use

./gpuowl -maxAlloc 3G -prp 96359411 -time

with GpuOwl VERSION v7.2-16-g1a50f11 from git, compiled with gcc 10.2.0, and it works, but produces incorrect results with ROCm 3.9 and my AMD Radeon R9 Fury X. Not sure why, previous ROCm version worked for me fine.

mandelbulber2 you can compile from source, or use the precompiled AppImage from their github release pages.

As of Blender OpenCL, ROCm 3.9 looks to have issues with it. GPU memory access fault in dmesg. Requires restarting the X server to recover from memory issues. kernel or ROCm bug. According to documentation on Blender website, OpenCL is maintained and is supported, so it should work.

Would be nice to try with ROCm 4.0 on 6800 XT.

[mibo]

A good old darktable test would also be nice...
...if you run OpenCL test the next time (6900XT?)

[GreenByte]

Michael, you can also add gpuowl and mandelbulber2 to the OpenCL tests. Both can be run from command line too. Maybe they are even in pts already. I don't know.

I.e. I use

./gpuowl -maxAlloc 3G -prp 96359411 -time

with GpuOwl VERSION v7.2-16-g1a50f11 from git, compiled with gcc 10.2.0, and it works, but produces incorrect results with ROCm 3.9 and my AMD Radeon R9 Fury X. Not sure why, previous ROCm version worked for me fine.

mandelbulber2 you can compile from source, or use the precompiled AppImage from their github release pages.

As of Blender OpenCL, ROCm 3.9 looks to have issues with it. GPU memory access fault in dmesg. Requires restarting the X server to recover from memory issues. kernel or ROCm bug. According to documentation on Blender website, OpenCL is maintained and is supported, so it should work.

Would be nice to try with ROCm 4.0 on 6800 XT.

Blender OpenCL works great with PAL (5700xt), so must be a ROCm issue (in fact, 5700xt can only do viewport tracing with ROCm, but that hopefully will change in 4.x series)

[bridgman]

We have Blender working pretty well over the ROCm back end as well as of the 20.45 driver. Not sure if all those changes made it into the 4.0 ROCm release but will be checking.

[cen1]

We are getting full, well working with good performance, day one open source graphics/compute support from AMD in year 2020 and all some supposed "linux enthusiasts" on this forum can comment is "whatever" and "wake me up when.." or "why don't they do X" and "it is still missing Z".

Just stop f* complaining for once and remember how the Linux GPU driver and gaming landscape looked 10 or 15 years ago and where we are today. God damn.. the nerve of some people..

/end rant

[oleid]

Looks good, but the memory is very slow and in some specific tasks perform very badly.
If you have a mixed use case like gaming + compute, it's a waste of money.

Is it? It all depends on your compute needs. The vulkan inference tests of the other article look quite nice. As well as these OpenCL tests (short of the artifical memory benchmarks). Don't forget these GPUs have more memory than what Nvidia has to offer today.

I can't wait to see how MiOpen will perform in ROCm 4.0.

[Michael]

We have Blender working pretty well over the ROCm back end as well as of the 20.45 driver. Not sure if all those changes made it into the 4.0 ROCm release but will be checking.

Looks like I got Blender OpenCL running now with the 20.45 driver.... So will at least have benchmarks there.

I tried the latest Folding@Home fahbench but that seems to be crashing, haven't looked into it any further yet to see if there is an easy workaround.

And oooo unrelated noticed today NAMD has now started putting out HIP/ROCm binaries that will hopefully work on RX 6800 when ROCm 4.0 is released.

[birdie]

Looking forward to full RDNA support in ROCm.

Looking forward to see how birdie is going to misrepresent this one.

Where and what have I misrepresented earlier? XFCE not supporting an anti-developer project called Wayland?

Heavy RTRT games not looking good on the RX 6000 series? Go check Control and Minecraft performance figures.

You're so funny.

Speaking of this review: I guess fewer than 1% of Phoronix readers have used OpenCL applications, so I don't understand how the review is relevant for most people here and why you personally are so excited. Have you participated in RDNA 2.0 development? Are you a major AMD shareholder? Then why? You're celebrating a company which is seemingly in cahoots with NVIDIA and sells its GPUs at very comparable prices? In the past it was called a collusion and it wasn't a reason for celebration. Times have changed I guess. Open Source drivers and shit and AMD is so great by selling your half a grand GPU whose cost price is at most $150.

Secondly, yeah, I'm gonna spoil your celebration here: GB5 shows that NVIDIA's OpenCL performance is quite good and also GB5 is not a single benchmark - it's a whole set of benchmarks (11 tests actually), so it's kinda unfair that its weight in the overall score is so low, while it must be multipled by at least five:

Sobel	276048 71.4 Gpixels/sec
Canny	173466 10.9 Gpixels/sec
Stereo Matching	734123 1.04 Tpixels/sec
Histogram Equalization	149461 26.4 Gpixels/sec
Gaussian Blur	191242 10.5 Gpixels/sec
Depth of Field	612038 7.10 Gpixels/sec
Face Detection	51063 393.1 images/sec
Horizon Detection	214424 5.29 Gpixels/sec
Feature Matching	50298 1.04 Gpixels/sec
Particle Physics	1089468 29018.9 FPS
SFFT	172478 2.38 Tflops

[birdie]

We are getting full, well working with good performance, day one open source graphics/compute support from AMD in year 2020 and all some supposed "linux enthusiasts" on this forum can comment is "whatever" and "wake me up when.." or "why don't they do X" and "it is still missing Z".

Just stop f* complaining for once and remember how the Linux GPU driver and gaming landscape looked 10 or 15 years ago and where we are today. God damn.. the nerve of some people..

/end rant

How many distros contain Linux 5.9 out of the box? Arch/Gentoo/LFS? They are not distros for normal people.

As I've said before a hundred times already, Linux kernel development model is bonkers as you either get stability or support for new hardware but never both. Drivers must be decoupled from the kernel to make Linux even remotely interesting as a desktop OS. Thanks to Google we'll soon have an open source kernel which works exactly like I expect: Zircon. I will gladly replace Linux with Zircon if NVIDIA starts supporting it. Don't wanna deal with regressions, bugs and GPL madness each three months.

[mppix]

...lots of b.s. ...
Heavy RTRT games not looking good on the RX 6000 series? Go check Control and Minecraft performance figures.

Yep, but AMD titles look much better (than Nvidia) so congratulations on the nitpicking for misreppresentation.

...lots of b.s. ...
Speaking of this review: I guess fewer than 1% of Phoronix readers have used OpenCL applications

I bet more than 1% use either libreoffice, blender, or darktable, ... just to name a few.

...lots of b.s. ...
Then why? You're celebrating a company which is seemingly in cahoots with NVIDIA and sells its GPUs at very comparable prices? In the past it was called a collusion and it wasn't a reason for celebration. Times have changed I guess.

The word that you are looking for is called competition. We all benefit from it. For example, Nvidia did not sell their high end (Ti and Titan) GPUs at reasonable prices until big navi happened.

.. delusional bla bla bla ..