Blender 3.3 AMD Radeon HIP vs. NVIDIA CUDA/OptiX Performance

[baka0815]

Michael
maybe you can organise a growdfunding for the 6950xt its the cheapest highend gpu price for many years.

That's a good idea, I think.

Did you consider crowdfunding for testing specific versions of GPUs, Michael?

[JellyBrain]

Michael
maybe you can organise a growdfunding for the 6950xt

I think he should just wait for RDNA 3, no need to invest that much money just before a new generation.

[qarium]

Michael

the 6950xt is down to 900€ https://geizhals.de/xfx-speedster-me...loc=at&hloc=de

according to userbenchmark the 6950xt is 27% faster than the 6800xt.. https://gpu.userbenchmark.com/Compar...4089vsm1843533

maybe you can organise a growdfunding for the 6950xt its the cheapest highend gpu price for many years.

i did see in the USA the same card is much cheaper than in europa/germany so maybe you get it for 800dollars..

also can you test the polaris support in the open source version of ROCm/HIP in blender 3.3 ?

[Grinness]

Michael


For curiosity I downloaded Blender 2.93.0 from:

https://download.blender.org/release/Blender2.93/

Cycles renderer does not have that Noise Threshold option in 2.93 (pre Cycles X) -- at least I could not find it under the Cycles settings.

Again I suspect that the Noise Threshold is a Cycle X feature -- and AFAIK CUDA backend uses Cycles X by default -- which would explain the difference in performance (CUDA vs HIP) between 3070 and 6800 across bmw27 and classroom scenes

Thanks

Michael

... I found a Noise Threshold option in 2.93.0 (under Adaptive Sampling check-box)

Loading the bmw27 file in Blender 2.93.0 the value for Noise Threshold is 0.00000 (adaptive sampling unchecked)
Loading the bmw27 file in Blender 3.3 the value for Noise Threshold is 0.01 (noise threshold unchecked)

Rendering using Blender 2.93.0, Cycles backed OpenCL (on ROCM 5.2.3 -- Feature Set supported) I get:

* no Adaptive sampling: 58 seconds
* Adaptive Sampling checked and Noise Treshold 0.00 (default value as read from file): 40 seconds
* Adaptive Sampling checked and Noise Treshold 0.01: 20 seconds


Using Blender 3,3, Cycles backed HIP (on ROCM 5.2.3 -- Feature Set supported) I get:

* no Noise Threshold: 22 seconds
* Noise Threshold 0.01 (default value as read from file): 13 seconds


I am not sure how the above plays with CUDA ... and I still cannot explain difference in performance (CUDA vs HIP) between 3070 and 6800 across bmw27 and classroom scenes cheking/uncheking that noise threshold

Thanks


EDIT: just to be pedantic, blender 3.3 projects have Noise Threshold checked as default with vaule 0.01

[Grinness]

Michael

For the records I did the same exercise (set/unset the Noise Threshold, value 0.01) in classroom.blend, and I get exactly the same results in both cases (again rx 6800, HIP on ROCM 5.2.3, arch aur repos): 40 seconds

3070 takes ~36 seconds (difference between 6800 and 3070 is about 10%)

Why the huge difference in performance in bmw27_gpu between 6800 and 3070 when Noise Threshold is unchecked (22 seconds vs 16 seconds -- about 38%)?

I would expect the opposite: if 3070 has overall better performance than 6800, that should show bigger advantage in classroom as the scene is more complex than bmw27

Thanks

EDIT: note that I refer to CUDA vs HIP performance -- Optix has the advantage of using the tensor cores, HIP does not use the RT on AMD cards

Michael


For curiosity I downloaded Blender 2.93.0 from:

https://download.blender.org/release/Blender2.93/

Cycles renderer does not have that Noise Threshold option in 2.93 (pre Cycles X) -- at least I could not find it under the Cycles settings.

Again I suspect that the Noise Threshold is a Cycle X feature -- and AFAIK CUDA backend uses Cycles X by default -- which would explain the difference in performance (CUDA vs HIP) between 3070 and 6800 across bmw27 and classroom scenes

Thanks

[Grinness]

Hi Michael,

cloud you please help me understand how you run the benchmarks?

I have a rx6800 reference card, running ROCM 5.2.3 from arch repos (zen kernel 5.19.10 updated to the latest version on arch repos).

When I open the bmw27_gpu.blend in Blender 3.3 I see that Cycles renderer on GPU is selected (HIP backend -- this is due to my default for cycles backend) , but 'feature set' is Experimental and blender shows a warning.
Switching 'Feature Set' to 'Supported' I get a render time of 21.97 seconds

However, the default settings for cycles in blender 3.3 are completely different from the settings loaded from the bmw27_gpu.blend
Default settings for blender 3.3. for Cycles rendering have Noise Threshold checked and at value 0.001 (my understanding is that the the Noise Threshold is a feature of Cycles X -- correct me if I am wrong)

If I check the Noise Threshold (value 0.01) in bmw27_gpu,blend (Feature set is 'supported', Cycles backed HIP -- all other parameters unchanged) rendering time is 13.10 seconds -- which is in line with 3070 ~ 16 secs

Is the cuda backend using Cycle X -- aka using that Noise Threshold?

Thanks


EDIT: fixed typos

Michael

For the records I did the same exercise (set/unset the Noise Threshold, value 0.01) in classroom.blend, and I get exactly the same results in both cases (again rx 6800, HIP on ROCM 5.2.3, arch aur repos): 40 seconds

3070 takes ~36 seconds (difference between 6800 and 3070 is about 10%)

Why the huge difference in performance in bmw27_gpu between 6800 and 3070 when Noise Threshold is unchecked (22 seconds vs 16 seconds -- about 38%)?

I would expect the opposite: if 3070 has overall better performance than 6800, that should show bigger advantage in classroom as the scene is more complex than bmw27

Thanks

EDIT: note that I refer to CUDA vs HIP performance -- Optix has the advantage of using the tensor cores, HIP does not use the RT on AMD cards

[Grinness]

Hi Michael,

cloud you please help me understand how you run the benchmarks?

I have a rx6800 reference card, running ROCM 5.2.3 from arch repos (zen kernel 5.19.10 updated to the latest version on arch repos).

When I open the bmw27_gpu.blend in Blender 3.3 I see that Cycles renderer on GPU is selected (HIP backend -- this is due to my default for cycles backend) , but 'feature set' is Experimental and blender shows a warning.
Switching 'Feature Set' to 'Supported' I get a render time of 21.97 seconds

However, the default settings for cycles in blender 3.3 are completely different from the settings loaded from the bmw27_gpu.blend
Default settings for blender 3.3. for Cycles rendering have Noise Threshold checked and at value 0.001 (my understanding is that the the Noise Threshold is a feature of Cycles X -- correct me if I am wrong)

If I check the Noise Threshold (value 0.01) in bmw27_gpu,blend (Feature set is 'supported', Cycles backed HIP -- all other parameters unchanged) rendering time is 13.10 seconds -- which is in line with 3070 ~ 16 secs

Is the cuda backend using Cycle X -- aka using that Noise Threshold?

Thanks


EDIT: fixed typos

[NeoMorpheus]

Nope never received any 6900/6950 from AMD, hence why it wasn't tested in this or any other articles... :/

That sucks and really weird that they never sent you either one.

I guess is time to fire up another tweet!

[RemcoL]

Can you perhaps add perf/watt graphs? Absolute performance is nice to be on top of benchmarks, but with all the constraints nowadays perf/watt is way more interesting. At least to me that is.

[ET3D]

It's not half bad for AMD, if you discount Optix. Shows that HIP has potential. We'll have to wait for Blender 3.5 to see how much ray tracing helps on the AMD side.