Karma Sink

Environment variables

Variable	Effect
`KARMA_XPU_OPTIX_CURVE_OPT_LEVEL = 1`	Memory consumption on OptiX devices can be reduction (with some performance loss) `1` = low-mem, `2` = mid-mem (default), `4` = high-mem
`KARMA_XPU_DEVICES=optix`	Use only Optix devices
`KARMA_XPU_DISABLE_OPTIX_DEVICE = 1`	Karma XPU will not use any Optix GPU devices
`KARMA_XPU_DISABLE_EMBREE_DEVICE = 1`	Karma XPU will not use any Embree CPU devices
`KARMA_XPU_DISABLE_DEVICE_n = 1`	Karma XPU will not use device number n. For example, to disable device 3 in a multi-GPU setup `KARMA_XPU_DISABLE_DEVICE_3 = 1`
`HOUDINI_TEXTURE_DISK_CACHE`	Set texture caching location "dir=D:/Temp/houdini_rat"
`KARMA_XPU_OPTIX_HQ_LINEAR_FILTERING`	Default OFF. Force Karma XPU Optix device to do high-quality linear filtering (which will come at a performance cost)
`KARMA_XPU_OPTIX_FORCE_CONSTANT_FOLDING`	Default OFF. This will force XPU to hardcode parameter values directly into shader code. It can improve performance of rendering by ~20%, but will mean shader compilation will be more frequent and take longer.

XPU Compiling

Compile cache location is C:\Users\Keerah\AppData\Local\NVIDIA\OptixCache\Houdini21.0

Sampling

Pre-compiling benefits interactivity and first pixel time
Adaptive sampling video

My notes on rendering Pyro with Karma

These checkpoints will make your render experience wayyyy moooore easier and snappy!

Use the Pyro Postprocess SOP node to cull the Vel volume by the Density, it saves a humongous amount of disk space (and performance) when caching volumes! Monitor the volume primitive sizes in info panel and you'll see that the Vel volume can be many folds bigger than the Density. Most of the time I also convert VDBs into 16 bit, all these options are present in that node.
When importing those volumetrics into Solaris always limit the Volume LOP explicitly to the volume primitives you really need to save (a lot of) VRAM. There is a field where you can specify the primitives. You might not need the Vel volume at all if you render with no motion blur.
USD cache all the things, this will bring your First Pixel time in hashtag#Karma to nearly zero! Especially when you have the "Render with single process" option on (then Karma caches the entire scene prior the rendering, and it can take pretty long)
I encountered a strange USD time sampling behavior when the volume was updating fine while inside Houdini but was frozen in the external renderer. A quick workaround is to USD cache it too. USD volume cache is very compact, since it does not contain the actual VDBs, but simply references them instead.

Motion Blur

Ultimate guide to motion blur in Houdini 22 using Solaris and Karma

Houdini 22 Ultimate Motion Blur Guide

Video Summary:

Key Techniques & Concepts

Velocity Blur: The standard, lightweight method for 95% of cases. It works by interpolating linearly between frames.
Acceleration Blur: A more advanced, efficient technique that provides curved motion blur. It requires at least three geometry samples to function correctly and is ideal for rotational objects like fans or wheels.
Interpolated Geometry/Deformation Blur: Uses a Retime node (set to cubic interpolation) and a Cache node to generate synthetic subframe data. While more accurate than acceleration blur, it can be computationally heavier.
Precise Subframe Data: The most accurate method involves using $FF (fractional frames) in expressions to generate real subframe animation data, which is then captured via a Cache node.
Particle Motion Blur: Limited to velocity or acceleration blur, as caching is impractical for particles due to changing point counts. The creator strongly advises against using particle trails because they fail when camera motion blur is introduced.

USD time samples filtering

A critical step when rendering to disk or Mplay is to change the Filter Time Samples setting in the USD Render ROP from always to never. This ensures that subframe data is not filtered out during the export process. If the stage is USD cached (with subsamples), then time filtering defaults to the cached sampling and this parameter can be always.

Recommendations

The creator recommends acceleration blur for general use due to its balance of speed and quality.
Reserve deformation blur (with caching) for 'hero' assets where extreme precision is required for visible motion curvature.

My benchmarks

XPU performance vs consumption

RTX 3080 + 3950X

XPU Device	temp limit, C	power limit, W	sample share, %	consumption, Wh
RTX 3080	65	-	78	250
Ryzen 3950X	-	-	22	190

RTX 3080 + RTX 2080

XPU Device	temp limit, C	power limit, W	sample share, %	consumption, Wh
RTX 3080	65	-	65	270
RTX 2080ti	65	-	35	120