October 27, 2026 Maddie Quach

Karma XPU Silent CPU Fallback: Causes, Checks and Fixes

Executive SummaryKey Takeaways: Karma XPU Silent CPU Fallback
Houdini Solaris / Karma XPU

What “silent CPU fallback” is, how to catch it and how to stop it:

  • What it is: when an OptiX GPU device fails, for example because it runs out of memory, SideFX’s documentation says the other devices, including the CPU, pick up its work. The frame finishes, just much slower.
  • Why it is easy to miss: the render does not stop. The device status (fail) shows in the viewport render stats, and husk saves render statistics into the EXR, but a farm job can still report success.
  • How to catch it: watch frame times, read the device list, and for farm jobs set KARMA_XPU_DISABLE_EMBREE_DEVICE so a failing GPU has no CPU to hide behind.
  • On iRender: RTX 5090 servers with 32 GB per GPU give a heavy scene more room, and you can open the render stats on the server and watch every device while the first frames render.

A Karma XPU sequence that renders in minutes per frame on your workstation suddenly takes ten times longer on another machine, yet the job log shows no error. In most cases one or more GPUs stopped contributing and the CPU kept going on its own. Artists call this the Karma XPU silent CPU fallback.

This guide explains what actually happens inside Karma XPU, the common causes, how to see it in the render stats and husk output, and how to make it fail loudly instead. It is based on the SideFX Houdini documentation and on SideFX staff answers on the SideFX forum.

What the silent CPU fallback really is

Karma XPU renders each frame with several devices at once: every NVIDIA GPU is an OptiX device and the CPU is an EmbreeCPU device. SideFX’s documentation describes what happens when one of them fails: if a device fails, as may be the case with a GPU running out of memory, the other devices, including the CPU, pick up the load. SideFX also says the result should be identical whatever mix of devices renders it.

That design keeps renders from crashing, but it has a cost. SideFX notes that the GPU devices contribute far more than EmbreeCPU, so when the GPUs drop out the CPU finishes the frame at CPU speed. The image looks right; only the time and the device list tell you something went wrong.

iRender
The silent CPU fallback
KARMA XPU
One render serverGPU 1OptiXrenderingGPU 2OptiXfail (OOM)GPU 3OptiXrenderingGPU 4OptiXrenderingEmbreeCPUpicks up the loadwork movesWhat you seeGPUs renderCPU finishes=normal frame timemuch slower, no errorSideFX: if a device fails, the other devices, including the CPU, pick up the load
OptiX GPUFailed deviceCPU device
From SideFX Karma XPU documentation

Three ways a Karma XPU render ends up on the CPU

What happensTypical causeWhat you see
The GPU never startsNVIDIA driver older than 570, a GPU below Compute Capability 5.0, or KARMA_XPU_DISABLE_OPTIX_DEVICE / a KARMA_XPU_DEVICES list without optix set somewhere in the environmentNo OptiX devices, or OptiX shown as disabled; EmbreeCPU does all the work from the first frame
The GPU fails partway throughThe scene, textures or AOV buffers no longer fit in that GPU’s memory; memory is not pooled across GPUsOptiX device status changes to fail, often on the heaviest frames only
The GPU keeps rendering, but slowlyGPU memory filled and Karma XPU started evicting texture tiles to keep goingNo device fails, but frame time rises; eviction shows in the render stats

From SideFX’s Karma XPU documentation. The last row is not a CPU fallback, but it looks similar from the outside, so check for it too.

How to catch it

iRender
What a silent CPU fallback looks like
per-frame time
timeframe12345678910GPU devices: fail(e.g. out of GPU memory)OptiX GPUs rendershort framesEmbreeCPU finishesevery frame: long times,job still “succeeds”✓ The image is the same, so only frame time and the device list show that the GPUs dropped outWatch both on the first frames of every job, or render GPU-only so a failure stops the job instead
Frames rendered with the GPUsFrames finished by the CPU devicePoint where the GPUs fail
Schematic, not a benchmark
  • Read the device list. In the viewport, the devices appear in the top right under the render progress, each with a status (idle, init, disabled or fail) and its share of the image. Turn on Render Stats in the Guides tab of Display Options for more detail, including why a device failed.
  • Keep the render statistics from batch renders. SideFX says husk saves Karma’s render statistics as metadata in the rendered EXR. The renderstatsreport utility turns that into an HTML report. SideFX notes that images from Karma XPU carry fewer statistics than Karma CPU, so treat the report as a check, not the only check.
  • Log husk output. Run husk with a higher verbosity level (for example –verbose 9, as farm submitters commonly do) and keep the text log with each frame range.
  • Watch frame times. A frame that takes several times longer than its neighbours with no change in the shot is the most common sign, as in the figure above.

Make a GPU failure stop the job instead

On a farm you usually want a broken frame to fail, not to finish slowly. SideFX documents two variables that remove the CPU device: KARMA_XPU_DISABLE_EMBREE_DEVICE, or KARMA_XPU_DEVICES=optix to use only the OptiX GPU engine. With no CPU device, a GPU failure has nothing to fall back to, so the problem shows up straight away.

Set the variable in the job’s environment, a Houdini package file or houdini.env on the render machines, render the heaviest frame, and only then send the full range. Remove it for interactive work if you like the CPU helping in the viewport.

iRender
Make failures visible: default vs GPU only
VARIABLES
DefaultCPU device includedGPU failsCPU finishes every framejob "succeeds", found lateGPU onlyKARMA_XPU_DISABLE_EMBREE_DEVICE=1STOPGPU failsnothing to fall back tojob stops at once: fix the causeOr set KARMA_XPU_DEVICES=optix; put it in the job environment, a package file or houdini.env
Frames on GPUsFrames on CPUGPU failure
From SideFX; bar heights are illustrative

Fix the cause

  • Driver and hardware: SideFX requires NVIDIA driver 570 or later and strongly recommends 575 or later on Windows; GPUs need Compute Capability 5.0 or higher.
  • Variables: search the environment, package files and houdini.env for KARMA_XPU variables left over from testing.
  • Memory: each GPU must hold the scene by itself. Lower texture resolution where it does not show, for heavy hair or fur try KARMA_XPU_OPTIX_CURVE_OPT_LEVEL=1 (SideFX’s low-memory setting), and remove AOVs you do not need. SideFX notes that with deep images and Karma XPU, husk has to allocate the whole raster, so deep output costs much more memory than flat EXRs.
  • Windows memory spill: a SideFX developer explained that on Windows, GPU memory can spill into system memory (“shared memory”) and still perform reasonably, but this does not work on Linux or with 2 GPUs, so do not count on it on multi-GPU render machines.
  • More memory per GPU: if the heaviest frame is close to 24 GB, a 32 GB RTX 5090 gives it room to stay on the GPU.
iRender
Fixing the cause
CHECKS
CHECKWhat to look forNVIDIA driver570+, 575+ recommended on WindowsCompute Capability5.0 or higherLeftover variablessearch for test KARMA_XPU variablesMemory per GPUeach GPU must hold the whole sceneHair and furKARMA_XPU_OPTIX_CURVE_OPT_LEVEL=1Check each device's status (idle, init, disabled, fail) in Render Statshusk saves render statistics as metadata in the EXR
Most commonCheckRequirement
From SideFX

SaaS or IaaS render farm when Karma XPU may fall back to the CPU?

Both farm models produce the same image, because Karma XPU gives the same result on any mix of devices. The difference is how quickly you find out that the GPUs dropped out, and what you can change when they do.

Your situationSaaS render farmIaaS render farm (iRender)
The scene is tested and fits well within GPU memoryA good fit. Upload the project and the farm renders the frames, often across many machines, with no server to manage.Also works, but you start, use and stop the server yourself.
You have a deadline and need to know when rendering starts and endsJobs join the farm’s queue, so the start time depends on how busy the farm is; many farms sell higher-priority plans to start soonerThe server runs only your job from the moment it boots, so you control when rendering starts and can time the whole job
You need to see whether every GPU is renderingYou read the job logs and frame times the farm gives you.Open the viewport render stats on the server and watch each OptiX device live.
A GPU fails on heavy frames and you need to find out whyEach test is a job: change the scene, submit, wait and read the log.Change textures, AOVs or curve settings on the server and render the same frame again straight away.
The heaviest frame is close to 24 GBCheck which GPUs the farm uses and how much memory each one has.Use RTX 5090 servers with 32 GB on every GPU.

A comparison of how the two farm models work for this task, not a speed test.

A simple rule: if the scene is tested, fits comfortably in GPU memory and you only need frames, a SaaS farm is a good fit; ask it to add the GPU-only variable so failures stop the job. If the heaviest frames are close to the memory limit, or you need to find out why a GPU fails, work on an IaaS server where you can see every device.

A fallback check on an iRender server

iRender
Works 1:1 like your own PC: we keep your setup, you just get more GPUs
IaaS render farm
Your workstationscene.c4d · .blend · .hip1 GPUYour iRender serverscene.c4d · .blend · .hipup to 8 GPUs=same setupWindows or Ubuntu desktopYour software, in the version you chooseYour plugins and scriptsYour licensesYour scene and render settings, unchanged1 GPUGPU power: the only difference1, 2, 4 or 8 × RTX 5090 / 4090✓ Only the GPU count changes: open your scene and render the way you do at home
Same on both machinesTip: upload the full project folder with relative paths so textures are found
  1. Upload the project and caches to iRender Drive before you start the server; uploading is free and does not use rental time.
  2. Start a 4x RTX 5090 server (32 GB on each GPU), install Houdini once and sign in with your own SideFX license. Billing is per minute.
  3. Open the heaviest frame in Solaris with Render Stats on and check that every OptiX device is rendering and none shows fail.
  4. Render it with husk and KARMA_XPU_DISABLE_EMBREE_DEVICE=1, with a verbose log, and note the frame time.
  5. Render the sequence and compare frame times with that first frame; if the deadline is tight, start more servers and give each one a frame range.

Before you render the sequence on another machine

  • Use the same Houdini release and an NVIDIA driver of version 570 or later.
  • Check the environment, packages and houdini.env for KARMA_XPU variables you did not mean to keep.
  • Make sure the heaviest frame fits in one GPU’s memory, with room for its AOVs.
  • Render the heaviest frame first with Render Stats or a verbose husk log, and confirm every OptiX device is working.
  • Keep the EXR render statistics and compare frame times across the sequence.

Frequently asked questions

1. What is the Karma XPU silent CPU fallback?

It is what happens when the OptiX GPU devices stop contributing, for example because a GPU runs out of memory. SideFX’s documentation says the other devices, including the CPU, pick up the load, so the frame still finishes, but at much lower speed and without stopping the job.

2. How do I know if Karma XPU is rendering on the CPU only?

Look at the device list in the viewport render stats: each device shows a status and its share of the image, and a failed GPU shows fail. For batch renders, keep a verbose husk log, check the render statistics saved in the EXR and compare frame times.

3. Does Karma XPU move data to system RAM when a GPU is full?

Textures are streamed to the GPU in tiles, and when GPU memory fills Karma XPU evicts non-essential memory and continues, which SideFX says slows the GPU down significantly. If the GPU still cannot cope, the device fails and the other devices finish the frame.

4. How do I make Karma XPU fail instead of falling back to the CPU?

Set KARMA_XPU_DISABLE_EMBREE_DEVICE, or KARMA_XPU_DEVICES=optix, in the environment of the render. Both remove the CPU device, so a failing GPU has nothing to fall back to and the problem shows up straight away.

5. Why does Karma XPU show no OptiX devices at all?

The most common causes are an NVIDIA driver older than version 570, a GPU below Compute Capability 5.0, or an environment variable such as KARMA_XPU_DISABLE_OPTIX_DEVICE set in the environment, a package file or houdini.env.

6. Can I check for CPU fallback on the machine that renders my final frames?

Yes, on an IaaS render farm. On iRender you rent the whole server, so you can open the heaviest frame in Solaris, watch every OptiX device in the render stats and then render the sequence with husk on the same GPUs.

Render your Houdini Karma XPU scenes on iRender

iRender is an IaaS GPU render farm: you rent a bare-metal server with 4x NVIDIA RTX 5090 (32 GB) GPUs, an AMD Ryzen™ Threadripper™ PRO CPU, 256 GB ECC RAM and a 2 TB NVMe SSD, connect by remote desktop and work on it like your own workstation, with full administrator access. You install Houdini and use your own SideFX license; your server image is saved between sessions, so Houdini, your packages and the Karma XPU kernel cache stay installed. For Karma XPU we recommend the 4-GPU servers.

Multi-GPU servers are billed per minute (1-GPU servers have a one-hour minimum), and sessions of 3 hours or more get 10% off. Our 4-GPU servers are liquid-cooled; in our own measurements the GPUs hold 60–70°C under sustained render load. Elite support is available 7 days a week by hotline, live chat, Skype, WhatsApp and email.

This video walks through rendering a Houdini Karma XPU scene on an iRender server in five steps:

New users get a 100% bonus on the first top-up (minimum $50). See the Karma XPU render farm page for server options, or download the iRender GPU app.

If Karma XPU renders are slow on your machine, iRender’s Karma XPU render farm gives you a dedicated remote server where you install Houdini, bring your own license and render on multiple RTX 5090 GPUs (for Karma XPU, 4 GPUs is usually the most efficient setup).

Accelerate your production pipeline with high-performance IaaS render farm services and dedicated bare-metal workstations powered by 4x NVIDIA GeForce RTX 5090 (32GB GDDR7), high-frequency AMD Ryzen™ Threadripper™ PRO processors, 256GB ECC RAM, and dedicated 2TB NVMe SSD storage tailored for your pipeline:
, , , , , , ,

Maddie Quach

Maddie Quach, from iRender's customer support team, has written more than 100 articles for the iRender blog on Maya, Houdini, Arnold, Karma XPU and Redshift.
Contact

iRENDER GPU

Render Farm Pricing
Quick Start
Multi-GPU Benchmarks
iRender GPU App
Blog
Documentation
The Render Farm
The Company
Terms & Conditions

INTEGRATIONS

Autodesk Maya
Blender
Cinema 4D
Houdini
Karma XPU
Lumion
KeyShot
Unreal Engine
Twinmotion
Redshift
Octane
And many more…

iRENDER TEAM

Elite support 7 days a week
SATURDAY – SUNDAY: 6:00 AM – 11:59 PM
(UTC+7)
Hotline: (+84) 912-785-500
Skype: iRender Support
Email: [email protected]
Address 1: 68 Circular Road #02-01, 049422, Singapore.
Address 2: No.22 Thanh Cong Street, Hanoi, Vietnam.

Contact
이 페이지는 한국어로도 제공됩니다. 한국어로 보기 →