[Eduardo G]

Yes, try to disable NUMA. We've done some tests on a AMD threadripper machine 1950X (16 core = 2 x 8 core) and with NUMA enabled the render times more than double when compared to the UMA option.

[Artistus]

Thank you guys for your help. I tried with NUMA disabled. Results are the same in all cases of QPI snooping configuration, except when QPI is set to Home Dir (Early Snoop disabled, CoD disabled) it made time twice as much over what was in previous modes testing - showed about 16m remaining so I canceled Benchwell after 20 seconds.

May that behavior be related to the fact that my CPUs are engineering samples with QHUZ code on their lids? Or maybe it's some pretty obscure little bug in V4 code related to Numa or QPI snooping mode set in UEFI? I even suspect some mistake in ASUS UEFI itself - at least there can be a chance of it's presence. Anyway, it's weird behavior regarding that Maxwell V3 is working flawlessly on any UEFI configuration tested here...

[dmeyer]

There is a bug in v4 that appears on dual-cpu systems. If AlphaCustom channel is activated, it is 10x or more slower. This channel is on by default in benchwell.

Try turning off that render channel and you will see render time improve on dual Xeon.

NL is aware of bug. I just discovered it myself yesterday.

[Artistus]

There is a bug in v4 that appears on dual-cpu systems. If AlphaCustom channel is activated, it is 10x or more slower. This channel is on by default in benchwell.

Try turning off that render channel and you will see render time improve on dual Xeon.

NL is aware of bug. I just discovered it myself yesterday.

Thanks man! I tested this out and disabled alpha custom channel - and voila! Rendering time became pretty much the same as in V3.2.1.5. But I'm using custom alpha channels in all my work, so until this bug is fixed by NL, I can't use V4 for work.

It's interesting if there is the same bug present on render farms with a lot of CPUs in servers. If so, it significantly decreases efficacy of render farms when using custom alphas, which I'm always using to optimize render times and to be able to extract different objects from render picture without hassle. Strange how it's been not fixed so far.

[dmeyer]

There is a bug in v4 that appears on dual-cpu systems. If AlphaCustom channel is activated, it is 10x or more slower. This channel is on by default in benchwell.

Try turning off that render channel and you will see render time improve on dual Xeon.

NL is aware of bug. I just discovered it myself yesterday.

Thanks man! I tested this out and disabled alpha custom channel - and voila! Rendering time became pretty much the same as in V3.2.1.5. But I'm using custom alpha channels in all my work, so until this bug is fixed by NL, I can't use V4 for work.

It's interesting if there is the same bug present on render farms with a lot of CPUs in servers. If so, it significantly decreases efficacy of render farms when using custom alphas, which I'm always using to optimize render times and to be able to extract different objects from render picture without hassle. Strange how it's been not fixed so far.

You could try using a custom defined object ID color instead.

[Artistus]

You could try using a custom defined object ID color instead.

Yes, but it seems like it does not allow to use it for extra sampling. And sometimes I render 5 different custom alphas at once to have quick and flexible options for postprocessing in Photoshop. Anyway, for me using Object ID channel instead of custom alpha delivers much more hassle in my work than just custom alphas.