Rendering Houdini on Cloud GPU: A Setup Walkthrough

It’s 9 p.m. and the Karma XPU shot you kicked off three hours ago still shows 300 frames left. Your workstation’s estimate just climbed to 26 hours, and dailies are due Monday morning. A cloud GPU machine running Houdini works exactly like a second workstation sitting next to yours. The only thing that changes the math is whether you set it up in the right order, so you are not paying for idle wait time.

Image Source: SideFX

When you rent a cloud GPU machine for Houdini, you get a full Windows or Ubuntu desktop reached over remote desktop, not a web form that swallows a .hip file and hands back frames later. You install Houdini yourself, open your scene exactly as you would at your desk, and work the viewport live if the job calls for it. Two things matter before you touch anything: the billing clock starts the moment the machine powers on, not when you click render, and the Houdini and renderer licenses running on that machine stay yours to manage.

Image Source: Side FX

Not every Houdini task benefits from more GPUs, so matching the package to the task is where the cost control actually happens.

• Lookdev and test renders: a single RTX 5090 32GB (package 3i) or a single RTX 4090 24GB (package 3S) if the scene comfortably fits in 24GB. One GPU is enough; you’re iterating, not finishing.
• Final render with Karma XPU or Redshift: a 4x RTX 5090 (package 5i) or 4x RTX 4090 (package 5S). Remember that VRAM is per card, not combined, so the ceiling for your scene is still a single card’s memory, multiplied by however many frames you can split across the four GPUs.
• Heavy CPU-bound simulation: stay on a GPU-light package. Extra GPUs do nothing for simulation and just add to the hourly rate.

Write down the exact Houdini and renderer version your project is pinned to before you do anything else. Gather caches, textures, HDAs, and any USD files the scene depends on, and repoint hardcoded paths to $HIP or $JOB so they resolve correctly once the project folder moves.

Upload everything through iRender GPU App while the machine is still powered off. Uploading after boot means you’re paying per-hour rates to watch a progress bar.

Install Houdini 22, matching the build your project needs. Install an NVIDIA driver that meets Karma XPU’s current requirement (driver 570 or newer, with 575 or newer recommended on Windows), and install any third-party renderer plugin builds (Redshift, Octane, V-Ray, Arnold) rebuilt for Houdini 22 if your pipeline uses one.

Sign in through SideFX Login or your studio’s license server, whichever your production uses. That license is yours to manage on this machine, same as it would be on a studio workstation.

Set your HDA paths and any custom environment variables in houdini.env, then save a Machine Image. The first time through this setup usually runs 15 to 30 minutes. Save the image here and that cost does not repeat.

Open Karma XPU in the viewport and check the device information in render stats. Cross-check it against nvidia-smi (Ubuntu) or Task Manager (Windows) while a render is actually running, since XPU can quietly drop to CPU-only without an obvious error.

If that happens, work through the usual suspects in order: the driver version is below what Karma XPU needs, the KARMA_XPU_DEVICES variable is set to something that excludes a card you expected to use, or a shader in the scene isn’t supported by XPU (VEX networks and some Principled Shader setups convert with mixed results, since XPU converts them to USDPreviewSurface rather than running them directly).

Render a single frame first and compare it against your workstation’s version of the same frame. 

For USD and Solaris-based scenes, render headless through husk rather than keeping the viewport open. For .hip-based scenes without a USD stage, hbatch or hython can drive the ROP network the same way. Both routes need the appropriate license (Houdini Engine or a full seat).

A husk command for a 300-frame sequence looks roughly like this:

husk -R karma -f 1001 -n 300 -o JOB/render/shot.F4.exr myscene.usd

Watch the first handful of frames come in and note the time per frame. That number, multiplied by frames remaining, is a far more honest estimate of what the job will cost than anything guessed up front.

Pull the finished EXR sequence back down through iRender GPU App. If you changed anything in the environment during this session (a new plugin, an updated HDA), save the Machine Image again before you disconnect.

iRender currently offers both an RTX 4090 (24GB GDDR6X) and an RTX 5090 (32GB GDDR7), so this is a decision you can actually test rather than guess at. When you rent a machine from us, you get an empty remote workstation. You install Houdini yourself and you license it with your own account. iRender does not provide, lend or sell Houdini or Houdini Engine licenses. As the sections above show, the answer follows from how you organise your jobs, so ask SideFX about your specific setup before you scale.

This differs from a SaaS render farm, where you upload a scene and the operator’s cluster handles the software side. Here the environment is yours, which brings one real advantage for this topic. You can set up the exact license configuration you are entitled to use, run your hython or husk commands on a single machine, and watch which licenses get checked out before you add a second node. Testing a batch pipeline on one RTX 4090 or RTX 5090 machine with a Threadripper PRO CPU costs far less than discovering a license gap on ten.

We are having promotion: you will receive 100% bonus for the first recharge. Check it out now!

Check out Quick Start for Houdini at iRender:

Maximum Speed – Absolute Freedom

1. Can Karma XPU use multiple GPUs on a cloud machine? 

Yes. Karma XPU picks up every compatible NVIDIA GPU it detects, but each card addresses its own VRAM separately rather than pooling memory across cards. Your scene still has to fit inside a single card’s memory, no matter how many GPUs the package has.

2. What NVIDIA driver does Karma XPU need in Houdini 22? 

SideFX’s current Karma XPU documentation calls for driver version 570 or newer, with 575 or newer recommended on Windows. Check the Karma XPU page again at install time, since driver guidance can change between point releases.

3. Should I run Houdini simulations on a multi-GPU cloud machine?

Usually not. FLIP, Pyro, Vellum, and RBD solvers run mainly on CPU and RAM, so a multi-GPU package adds cost without speeding up the simulation itself. Simulate on a single-GPU package, then move to a multi-GPU package for the render pass.