Houdini Render Farm for Karma XPU, Redshift & Octane: Rent 2–8x RTX 5090 Servers

• A whole multi-GPU workstation, not a slot in a queue. Rent up to 8x RTX 5090 with 32 GB VRAM each. Every server has 256 GB ECC RAM and a 2 TB NVMe SSD, and the 4x and 8x servers run an AMD Threadripper PRO 5975WX with 32 cores and 128 PCIe lanes.
• Your Houdini build, unchanged. Install the exact production or daily build through the SideFX Launcher, sign in with your SideFX account or point to your license server, and load your HDAs, packages and OCIO config. Nothing on the farm side converts your .hip or USD files.
• Three GPU renderers on one machine. Karma XPU, Redshift and Octane all use every GPU in the server. Karma XPU runs best on 4 GPUs, while Redshift and Octane keep gaining speed up to 8. Karma CPU and Mantra jobs run on the same 32 CPU cores.
• Simulate and render in one place. Cook FLIP, Pyro or Vellum on the CPU, write caches to the local NVMe drive, then render them on the GPUs without moving terabytes between machines.
• You pay for render time, not for setup. Upload to iRender Drive for free before the server starts. Your first top-up of $50 or more is doubled, which covers installing Houdini and test frames. Multi-GPU plans bill per minute, and sessions of 3 hours or more get 10% off the whole bill.

Most Houdini render farms are submission farms: you upload a .hip or USD file and the farm renders it with its own Houdini build and plugins. iRender works differently. You rent the whole server and drive it yourself, so your build, HDAs and renderer versions stay exactly as they are.

Heavy Houdini shots usually fail for the same few reasons: the volume does not fit in VRAM, the cache does not fit on the disk, an HDA or package is missing, or the CPU becomes the bottleneck while cooking and exporting. The table matches each one to what an iRender server gives you.

The figure gives the short version. The table below goes row by row: how each heavy-scene problem shows up in production, and the exact hardware or access that handles it. If your shot hits more than one row, a 4x or 8x RTX 5090 server (plan 5i or 9i) covers all of them on one machine.

Every plan is a whole server. Pick the GPU count for your deadline and the card for your scene size: RTX 5090 when the scene needs more than 24 GB, RTX 4090 when it fits. Heavy simulations are best on the 4x and 8x servers, which have the 32-core CPU. For Karma XPU, a 4x server is the sweet spot; choose 8x for Redshift or Octane.

None of the three GPU renderers adds VRAM across cards. SideFX notes that in Karma XPU each GPU device is addressed separately and memory is not pooled, and Redshift and Octane also keep a copy of the scene on every GPU. The limit is 24 GB on RTX 4090 plans and 32 GB on RTX 5090 plans, however many GPUs the server has. More GPUs make each frame faster; they do not make room for a bigger volume.

• Karma XPU: if a GPU device fails, for example because it runs out of memory, SideFX says the other devices, including the CPU, pick up the load. The frame still finishes, but much more slowly. Check the device list in the render stats, or set KARMA_XPU_DISABLE_EMBREE_DEVICE=1 so a GPU failure stops the render instead of hiding on the CPU.
• Redshift and Octane: both have out-of-core options that keep part of the scene in system RAM when it does not fit in VRAM, at a cost in speed. Every iRender server has 256 GB ECC RAM for this.

If your volumes are close to 24 GB, choose an RTX 5090 plan.

• Karma XPU and Karma CPU: ship with Houdini. Karma XPU uses the CPU and every compatible NVIDIA GPU together. In our experience it runs best on 4 GPUs: an 8-GPU server works, but adds little speed, so choose a 4x server (5S or 5i) for Karma XPU. When you render with husk, SideFX checks out a Karma license, so sign in with the SideFX account or license server that holds your Houdini and Karma licenses. Rendering mostly with Karma? See our Karma XPU render farm.
• Redshift for Houdini: install the Redshift build that matches your Houdini build and sign in with your own Maxon license, one per server. Measured Redshift numbers are on the Cinema 4D Redshift render farm page.
• Octane for Houdini: install the Octane for Houdini plugin and sign in with your OTOY account. An Octane Studio+ subscription has no GPU limit, so it uses every card. See the Octane render farm page for benchmarks.
• Houdini Apprentice and Education: follow SideFX terms; for paid client work, use a commercial Houdini license.

These iRender videos record complete Houdini sessions, from uploading the project to downloading the frames. Render speed depends far more on the renderer and the shot than on the host application, so for measured frame times and costs see the Redshift and Octane benchmarks linked above.

Watch the full 8x RTX 4090 Redshift session for Houdini, from uploading the project to shutting the server down.

Submission farms are popular for long sequences because they split frames across many nodes. You can do the same on iRender: rent two, three or more servers, open the same project from iRender Drive on each, and give each server its own frame range. Billing is per minute per server, so the total cost stays close to rendering on one server, plus a few minutes of setup on each.

Each server opens the same project from iRender Drive and renders its own frame range, set in the Valid Frame Range of the ROP, or with husk on the command line, for example husk -f 1 -n 40 shot.usd for frames 1 to 40. Each server needs its own render license. Here is the math for a 120-frame example.

1. Register and top up. Your first top-up of $50 or more is doubled, so setup time is covered.
2. Upload your project and caches to iRender Drive. Transfer and storage are free, and you can upload before the server starts.
3. Start a server and connect over remote desktop. Install Houdini with the SideFX Launcher, sign in to your licenses, install your renderer plugin, and copy the project from Z: to the local C: drive.
4. Render with a ROP, husk or TOPs, copy the frames to Y: to download them, and shut the server down; billing stops.

Connect to your server from whatever machine you work on. The iRender GPU app runs on Windows, macOS (Apple Silicon and Intel) and Ubuntu/Debian, and the web app opens in any modern browser with nothing to install, so a Mac user can drive an 8x RTX 5090 Windows server.

Once connected, the server behaves like a local workstation. The quickstart below shows the full flow on iRender, from uploading the project to shutting the server down.

A multi-hour 8x GPU render needs the building around it to stay up, and studio work under NDA needs a provider that can show its security controls. iRender servers run in an Uptime Institute Tier III data center with ISO-certified operations.

Support is available 24/7 on weekdays and from 6:00 AM to 11:59 PM (UTC+7) on weekends, by hotline, Skype and email.

Hand-picked guides from the iRender blog for the problems Houdini artists ask us about most: licenses, caching, multi-GPU Karma XPU and crashes.