Bucket and Progressive Samplers' Analysis

[dean_dmoo]

Thanks, nice to see a clear overview!

Are you guys planning on unifying the render defaults in the next update? (Max AA and noise) Or is there a technical reason why progressive has higher default settings?

[jacksc02]

Nice clear explanation.
Thanks Lele!

[jasonhuang1115]

Too late to the party! Thanks for the great write-up, Lele!

[Micha]

Hi ^Lele^

last I observed a much lower GPU power usage during progressive rendering. The attached GPU-Z screenshot shows a progressive rendering at approx. 50% GPU load and a bucket rendering of the same scene at 98%. The render times of the bucket mode was halfen.

I use a five years old DualXeon E6 2687W 0 and two 2080ti. I ask me, is the reason a general limit of the progressive sampling or is it a hardware bottle neck?

-Micha

[^Lele^]

Micha, i think it's a question for the GPU team to answer.
I doubt it's an hardware thing, it looks like they made good advancements in the bucket renderer, but i am not current (with tests).

[kosso_olli]

I had the same issues for the progressive sampler on CPU as well, especially when rendering very small areas as crop or region. Setting the ray bundle size to a very low value helped solving this.

[Joelaff]

I had the same issues for the progressive sampler on CPU as well, especially when rendering very small areas as crop or region. Setting the ray bundle size to a very low value helped solving this.

You mean lowering the ray bundle size improved the cpu usage during progressive? Did it improve render time? Interesting. I had always assumed a higher bundle size would be more efficient. I guess I had better do some tests.

[^Lele^]

Fair warning: usage != speed.
Lowering RBS may just initiate more data transfers across the system's buses.
Ensure you benchmark the rendertime.

Also, there is likely a specific threshold for each CPU, and scene, below which things worsen, even considerably.
Mileage is going to vary quite a bit.

[Joelaff]

Precisely my point about asking if it improved render time.

Just because the cpu is doing more doesn’t mean rendering faster.

Why does progressive seem to have much more idle time on the cpu cores? Would be cool if that could improve. I think now that bucket has the automatic subdivision that fixes last bucket syndrome it is usually faster. But, with everything, YMMV

[^Lele^]

Why does progressive seem to have much more idle time on the cpu cores?

It doesn't for me, generally.
It does if the progressive caustics engine is on, due to the time it takes to gather and make sense of the caustics' data between passes.

[Joelaff]

It doesn't for me, generally.
It does if the progressive caustics engine is on, due to the time it takes to gather and make sense of the caustics' data between passes.

No caustics other than the default GI based caustics.

On the entire farm when using Progressive I frequently see the CPU usage well below 100% Sometimes down in the single-digit percentages for periods between pass updates I think. So I end up having to run multiple instances of Max to maximize usage. With Bucket it is almost always 100% these days (thanks to the Last Bucket Syndrome fix!).