Karma XPU is a selectable engine of the Karma renderer that makes use of both CPU and GPU resources simultaneously. It will use any compatible devices it can detect, including multiple GPUs. If a device fails, as may be the case with a GPU running out of memory, the other devices (including the CPU) will pick up the load and finish rendering the frame. No matter what mixture of devices it uses, the XPU engine should produce the exact same result.

Karma XPU is split into the OptiX GPU and Embree CPU engines. In systems with multiple graphic cards, you will also see an appropriate number of OptiX devices in the viewport. Note that each GPU device is addressed separately and memory is not pooled.

An important concept with Karma is sampling, because it has strong impact on render times and the quality of the rendered image. You can read more about sampling and how it works in Karma XPU in the Karma User Guide.

Because Karma XPU uses GPU hardware acceleration, it is very fast compared to the CPU-only engine. However, XPU may have less functionality, and/or generate different output from the CPU-only engine. Karma XPU is not a replacement for the Karma CPU engine, but rather a high-performance alternative with limited feature set.

• Karma XPU currently only supports CPU and NVIDIA GPU hardware (support for other GPU hardware will come in the future).

• NVIDIA users require driver version 570 or later (575+ strongly recommended for Windows; 580+ strongly recommended for Maxwell, Pascal, and Volta GPUs on both Windows and Linux), and a GPU with Compute Capability 5.0 (Maxwell) or higher.

• Machines without an NVIDIA GPU (e.g. Apple Silicon) will only make use of the Embree CPU device.

The Karma Material Builder is recommended as the way to build materials for Karma XPU. This will provide access to all of the compatible Karma and MaterialX nodes. If you need a strictly MaterialX-only set of shaders, use the USD MaterialX Builder. See the Materials chapter in the Karma User Guide for more details.

On startup, Karma XPU must compile shaders into device-specific formats. This process will cause the XPU to pause/stutter when you load a new scene or enable a new feature. The compiled results are cached, so rendering the same scene will not pause for shader compilation after the first time. When first using XPU, or with new installations of Houdini, or on a new GPU driver, the initial compilation steps can take up to 1-2 minutes. But because of the caching, further compilation stages should be a lot faster. During compilation, corresponding information is displayed in the top-right corner of the viewport.

In the viewport you’ll see the active devices in the top right of the screen under the render progress readout. The devices will be listed as either “OptiX” or “EmbreeCPU” with a status text in the square brackets. Status can be one of the following:

• idle - the device is not doing any work

• init - the device is initializing

• disabled - the device has been disabled

• fail - the device has failed

• a percentage number representing how much that particular device has contributed to the current image so far (its expected GPU devices like Optix will contribute far more than EmbreeCPU given Karma XPU has been written and optimized for GPU).

It is also possible to see more detailed information about the individual devices in the render stats.

When OptiX is compiling, corresponding stats are also displayed in the top-right. Compiled Render Kernel Cache out of Date indicates that Karma XPU pre-compiling could be used to reduce further compilation overhead.

• Karma XPU supports the following render settings:

  • Cancel Render on No Working GPU Devices

  • Path Traced Samples

  • Russian Roulette Cutoff Depth

  • Light Sampling Mode

  • Light Sampling Quality

  • Image Blur

  • IPR Downsample Factor (controls resolution of the preview frame)

  • IPR Reserve Threads (XPU is more responsive if you set this to 2 or 3)

  • Enable Denoising

  • Disable Motion Blur

  • Disable Depth of Field

  • Ray Bias

  • Automatic Ray Bias

  • Constrain by Maximum Roughness

  • Color Limit

  • Shared Color Limit

  • Indirect Color Limit

  • Automatic Headlight Creation

  • Dicing Camera

  • Offscreen Quality

  • Dicing Quality Scale

  • IPR Continuous Dicing

  • Pixel Filter Size (only Gauss pixel filter supported ATM)

  • Pixel Oracle (ie “Variance Oracle” and associated controls)

  • Random Seed

• Supports the following render geometry settings:

  • Enable Motion Blur

  • Motion Samples from Stage

  • Geometry Time Samples

  • Transform Time Samples

  • Motion Blur Style

  • Instance Motion Samples

  • Volume Velocity Blur Scale

  • Volume Velocity Lattice Quality (only for level sets)

  • Volume Step Rate

  • Volume Shadow Step Rate

  • Diffuse Limit (range 0 to 255)

  • Reflection Limit (range 0 to 255)

  • Refraction Limit (range 0 to 255)

  • Volume Limit (range 0 to 65535)

  • SSS Limit (range 0 to 255)

  • Limit Components

  • Limit Color

  • Uniform Volume (see Uniform Volume section below)

  • Render Visibility

  • Stencil Map (only supports Polymesh)

  • Stencil Threshold (only supports Polymesh)

  • Holdout Mode

  • Fix Shadow Terminator

  • SSS Traceset (maximum of 4 tracesets per scene)

  • LPE Tag

  • Enable Internal Reflection

  • Thinwall at Refract Limit

  • Direct Refraction Subset

  • Dielectric Priority

  • Enable Caustics

  • Caustics Roughness Clamp

  • Evaluate BSDF on Fake Caustics

  • Fake Caustics Color

  • Fake Caustics Opacity

  • Fake Caustics Fresnel Darken

  • Efficient Emissive Sampling (used for geometry lights, see Lighting section below)

  • Is Portal (see Lighting section below)

  • Portal Dome Lights (see Lighting section below)

  • Render Points As

  • Cull Backface (only supports Polymesh and PointMesh)

  • Dicing Quality

  • True Displacements (only supports MtlX Standard Surface VOP and MtlX OpenPBR Surface VOP with float/scalar displacement)

  • Dicing Minimum Depth

  • Dicing Maximum Depth

• Supports selection, highlight, and USD purposes when used as an interactive viewer.

• Supports motion blur (and all its associated settings), including light blur, camera blur, transformation blur, deformation blur and volume motion blur (Lookup Offset mode only).

• Supports all camera settings (including Physical Lens Shader).

• Supports USD Mesh/Polygon prims

• Supports USD BasisCurves prims, details:

  • Supports only linear, cubic-bspline, catmull-rom and cubic-bezier types

  • Supports only rounded-tubes rendering

  • Does not support disabling backface culling

  • Memory consumption on OptiX devices can be reduced (with some performance loss) via the KARMA_XPU_OPTIX_CURVE_OPT_LEVEL environment variable (1 = low-mem, 2 = mid-mem (default), 4 = high-mem).

  • S coord is supported (to achieve a cylindrical mapping), but requires N attribute (equivalent to ribbon normal) to serve as guiding orientation.

• Supports USD Points prims, details:

  • Supports disk, sphere and oriented point types (via the Render Points As geometry property)

• Supports USD Volumes, details:

  • Only XPU Pyro Preview VOP, Karma Volume VOP and Karma Whitewater VOP working

  • Does not support volume filter and width parameters.

  • Direct rendering of VDB points and ISO surfaces

• Supports Uniform Volumes, details:

  • Enable via Uniform Volume render geometry setting

  • Only supported on USD Mesh/Polygon and Points prims

  • Only Karma Volume VOP and Karma Whitewater VOP working

• Supports Gaussian Splats (alpha)

  • Import via the Bake GSplat SOP

  • Will intersect and “combine” correctly with other scene elements (as well as with other gsplats)

  • Can cast shadow into the scene (via a setting on the Bake GSplat SOP)

  • Does not support recieving lighting (or re-lighting)

  • Does not support LPEs

• Supports Nested Instancing (including per-instance properties)

• Supports Cryptomatte.

• Supports Deep Output

• Supports Named coordinate systems provided by UsdShadeCoordSysAPI

  • Supported in the MtlxPosition, MtlxNormal, MtlxTangent and MtlxBitangent nodes

  • Supported in the MtlxTransformPoint, MtlxTransformNormal and MtlxTransformVector nodes

• Supports Karma Sky Atmosphere LOP.

Set one or more of the following environment variables in your environment, json package, or houdini.env.