brians
KARMA_XPU_THREAD_COMPILE=1

- no powersave mode. It's set to ultimate performance
- all core light up with embree only (also with all devices) when rendering
- update, no difference
- KARMA_XPU_THREAD_COMPILE=1, all cores light up, but each of those processes don't seem to have any effect on the progress (maybe thats expected)
- I'll update with that info once I have it, but going in about 40 labs, this seems to be as slow as before.



I also gave this a spin on a 32 core threadripper with 2x 5090, the load is about double (10-20%), but then again, it has only half the cores (so I guess its about the same speed in the end).

Also tried it on an older machine with dual 20core xeons and a 3080. Same thing.

These are vastly different machines. The only thing they share is windows 11 and 32cores+. And they all work fine during normal rendering.

hmmmm... very strange.
and if you let one of threadripper machines finish, what is the total time? 2 hours? 20 hours?
thanks

brians
hmmmm... very strange.
and if you let one of threadripper machines finish, what is the total time? 2 hours? 20 hours?
thanks

it took almost 6 hours. This is a little bit faster than on the 64 core machine, but I think this is because the single thread performance on this machine is better.

I`m still experiencing same issues after the "fixed" build, eta was 5-6h for single GPU, i gave up after 3h
5090 + RTX PRO 6000, Threadripper 7980x W11, but same issue on different machine with Threadripper PRO 9985WX and single RTX 6000.

While I'm a fan of making things faster, I feel that compiling things I'm using just-in-time is probably absolutely fine for my home usage, considering my Mtlx library coverage in my average shader is fairly small.

In the studio however, doing this for our render farm might save quite a bit more time. Has there been much more thought to allowing us to customize the destination of these cache files to shared network disk (using lock-files/etc) so we can do can accumulate compiled code, and/or do a complete precompile upon installation/driver change? Perhaps as a cron which would mostly just NOP until it detected a change.

In short, I'll re-run on a (medium-grade: 32-core Threadripper) workstation and report the performance instead.

It took 55 mins. It might be nice to have an estimated total (eLapsed + Left), and a summary at the end?

jason_iversen
In the studio however, doing this for our render farm might save quite a bit more time

Hello! I think precompile is only useful in an interactive session, not in a render farm context.
regards,
F.

EbenF

jason_iversen
In the studio however, doing this for our render farm might save quite a bit more time

Hello! I think precompile is only useful in an interactive session, not in a render farm context.
regards,
F.

Hiya,

Certainly one could argue Karma shaders still compile quite quickly and might not be worth a shared-cache/precompiling on the farm - especially when rendering hero characters and FX - but when you're talking about hundreds or thousands of materials in larger layout scenes, it might help to save some minutes per-frame across multiple frames on a farm.

FYI, we already successfully cache shader compilations for our other renderer (Manuka) on our farm, where big shaders can sometimes take (many) minutes to compile and optimize. So this is area is pretty well understood for us already and it would hopefully be quite trivial to add Karma alongside our preexisting shader cache system.