Blender 3.4 HIP Performance With Radeon RX 7900 Series + RDNA3 OpenCL Compute Benchmarks

[WannaBeOCer]

That is an 'orange to apple' comparison

OptiX uses tensor cores, HIP does not use ray tracing hardware on AMD. HIP RT in blender 3.5 is supposed to 'bridge the gap' from a software enablement point of view.
To the best of my knowledge ROCm still does not use RT hardware

OptiX uses RT cores to accelerate ray tracing and uses an AI accelerated by Tensor cores to denoise the scene.

Hopefully AMD updates HIP-RT to utilize their new AI Accelerators in RDNA3 for denoising. They currently have an AI in Radeon ProRender for denoising.

https://gpuopen.com/radeon-prorender-suite/

Uses AMD’s MIOpen machine learning library or Microsoft® DirectML for AI-accelerated denoising and upscaling.

Nvidia’s Turing was ground breaking, glad to see AMD’s consumer GPUs are at feature parity with Turing. Gamers weren’t impressed with Turing but regarding content creation/compute it was leaps and bounds faster than Pascal.

[WannaBeOCer]

Michael

I just tested another of the sample blender files you report performance on (Fishy Cat -- which is a Blender 2.74 blend file) from:

https://www.blender.org/download/demo-files/

on AMD rx 6800 ROCm 5.3.0 render time is 11.11 seconds (this is the second time rendering, first time is 16.11 seconds)-- you report 43.26 seconds for same hardware, see first plot on:

https://www.phoronix.com/review/rx7900-blender-opencl/3

Steps to reproduce:

1. download blender zip
2. open file 'fishy_cat.blend'
3. select 'GPU Compute' device
4. hit 'F12'

The above results is in line with the OptiX results you report on 3080/3080Ti

Are we sure that the results in your article make sense?

EDIT: the 'readme' in the blend file explain:



The results I report above are after the final 'composite pops out'

EDIT2:
1. I am on blender 3.4 (arch linux)
2. I repeated the 'step to reproduce' process 3 time just to make sure -- obtaining equivalent results (best render time is 11.0 seconds)

Most reviewers increase the output resolution of a project to properly load the GPU. If they don’t they’ll end up with times similar to yours. Leading to similar results across all GPUs, making the results useless for consumers who actually need GPUs for rendering.

[Mushoz]

Haven't had a chance to look into it yet but it's been on my todo list.

If it doesn't end up working, please do let us know as well! So we at least know you tried, and we have confirmation it does not work.

[Grinness]

Most reviewers increase the output resolution of a project to properly load the GPU. If they don’t they’ll end up with times similar to yours. Leading to similar results across all GPUs, making the results useless for consumers who actually need GPUs for rendering.

Most reviewers provide information about the settings used if different from 'default'.
I can reproduce rx 6800 HIP render times on classroom (same results with & without the 'Noise Threshold' flag set), I can reproduce render time on bmw (without 'Noise Tresholds', with noise threshold I obtain 10 seconds faster render time).
To reproduce I simply open the blend file, set backend to HIP and Cycles to use the GPU

I haven't tested all the blender files Phoronix reports results on -- and I do not have nvidia hardware to cross check anyway.
Furthermore, it is not clear what 'default setting' is when it comes to blender files from 2.x era that did not have implemented the backends for which performance are reported.

The issue at hand is the credibility of Phoronix tests when it comes to comparison between different hardware, different backends, different scenes (blend files)

Phoronix praises itself for the reproducibility of metrics reported, see:
https://www.phoronix-test-suite.com/...p?k=phoromatic

Unfortunately in the present case it fails to convince that what is reported is actually correct.
This casts serious doubts also on its general methodology in bechmarking

Note that I raised already some of my concerns when the article on performance of Blender 3.3 on HIP vs CUDA/OptiX was published.
There I found that there was a lack of logical consistency of results for rx 6800 and RTX 3070 between classroom and bmw (vs 'Noise Threshold' enabled/disabled)

No answer was given at the time
Should I expect no answer to concerns this time too? -- I can be wrong in my verification of metrics, but please show me

[WannaBeOCer]

Most reviewers provide information about the settings used if different from 'default'.
I can reproduce rx 6800 HIP render times on classroom (same results with & without the 'Noise Threshold' flag set), I can reproduce render time on bmw (without 'Noise Tresholds', with noise threshold I obtain 10 seconds faster render time).
To reproduce I simply open the blend file, set backend to HIP and Cycles to use the GPU

I haven't tested all the blender files Phoronix reports results on -- and I do not have nvidia hardware to cross check anyway.
Furthermore, it is not clear what 'default setting' is when it comes to blender files from 2.x era that did not have implemented the backends for which performance are reported.

The issue at hand is the credibility of Phoronix tests when it comes to comparison between different hardware, different backends, different scenes (blend files)

Phoronix praises itself for the reproducibility of metrics reported, see:
https://www.phoronix-test-suite.com/...p?k=phoromatic

Unfortunately in the present case it fails to convince that what is reported is actually correct.
This casts serious doubts also on its general methodology in bechmarking

Note that I raised already some of my concerns when the article on performance of Blender 3.3 on HIP vs CUDA/OptiX was published.
There I found that there was a lack of logical consistency of results for rx 6800 and RTX 3070 between classroom and bmw (vs 'Noise Threshold' enabled/disabled)

No answer was given at the time
Should I expect no answer to concerns this time too? -- I can be wrong in my verification of metrics, but please show me

Don't you think it's strange a RX 5700 XT performed the same as your RX 6800? https://www.phoronix.com/forums/foru...83#post1362783

I suggest you compare against TechGage since they list the project output resolution: https://techgage.com/article/amd-rad...reator-review/

[finalzone]

Michael
Here is a sample of Blender benchmark using Sapphire Nitro+ Pure RX 6950 XT: https://openbenchmarking.org/result/...NE-SAPPHIRE603
Note system is a Ryzen 9 5900X.
Luxcorerender 2.6: https://openbenchmarking.org/result/...NE-6950XTLUX63
DSLC: http://localhost:8940/result/6950xt-...-dlsc-20221216

[smitty3268]

Michael

I just tested another of the sample blender files you report performance on (Fishy Cat -- which is a Blender 2.74 blend file) from:

https://www.blender.org/download/demo-files/

I took a look at the PTS source code, and i'm not very familiar with either it or blender, but it looks like it's getting the fishy cat and other test files from here: http://download.blender.org/demo/tes...k_20160228.zip

And that does appear to be a slightly different file size than the one you pointed out, so maybe there are differences there.

I also see these arguments something like this being passed to blender: blend -noaudio --enable-autoexec -b fishy_cat_gpu.blend -o output.test -x 1 -F JPEG -f 1

That's from a blender 3.2 test profile, i don't see anything for 3.4, but i'd guess it's probably running the same thing.

[Grinness]

Don't you think it's strange a RX 5700 XT performed the same as your RX 6800? https://www.phoronix.com/forums/foru...83#post1362783

I suggest you compare against TechGage since they list the project output resolution: https://techgage.com/article/amd-rad...reator-review/

I already checked results on TechGage, although those number originates from a different OS, driver etc, I can reproduce

With regards to the render time of a 5700XT on the fishy-cat blender scene from Phoronix reader -- I am not surprised. That scene is so simple that any modern GPU should show very comparable render times

[Grinness]

I took a look at the PTS source code, and i'm not very familiar with either it or blender, but it looks like it's getting the fishy cat and other test files from here: http://download.blender.org/demo/tes...k_20160228.zip

And that does appear to be a slightly different file size than the one you pointed out, so maybe there are differences there.

I also see these arguments something like this being passed to blender: blend -noaudio --enable-autoexec -b fishy_cat_gpu.blend -o output.test -x 1 -F JPEG -f 1

That's from a blender 3.2 test profile, i don't see anything for 3.4, but i'd guess it's probably running the same thing.

Ok, that may explain -- if the source files do not coincide -- I'll take a look

thanks

[Grinness]

I took a look at the PTS source code, and i'm not very familiar with either it or blender, but it looks like it's getting the fishy cat and other test files from here: http://download.blender.org/demo/tes...k_20160228.zip

And that does appear to be a slightly different file size than the one you pointed out, so maybe there are differences there.

I also see these arguments something like this being passed to blender: blend -noaudio --enable-autoexec -b fishy_cat_gpu.blend -o output.test -x 1 -F JPEG -f 1

That's from a blender 3.2 test profile, i don't see anything for 3.4, but i'd guess it's probably running the same thing.

Alright,

testing the fishy_cat_gpu.blend in:
http://download.blender.org/demo/tes...k_20160228.zip

I get similar results as in Phoronix , with rx6800, HIP, ROCm 5.3.0 :
43.96 (mine) vs 43.26 (Phoronix)
(note that enabling 'Noise Threshold' render time goes to about 25 seconds)

However, if I take classroom_gpu.blend ( or classroom_cpu.blend) in
http://download.blender.org/demo/tes...k_20160228.zip

the scene looks completely different from the file in:
https://www.blender.org/download/demo-files/

The room is empty and has no furniture or heaters.

interestingly the render times are not comparable anymore:
37.70 (mine) vs 49.86 (Phoronix)
(enabling 'Noise Threshold' does not produce any meaningful speed up in render time)

If I use the classroom_gpu.blend from:
https://www.blender.org/download/demo-files/

now render times are comparable (sort of)
46.84 (mine) vs 49.86 (Phoronix)
(enabling 'Noise Threshold' does not produce any meaningful speed up in render time)

Extremely confusing and not clear what has been tested