Pyro Sim Dies Halfway Through: Voxel Resolution, Substeps and Memory
A Pyro simulation that dies halfway has usually grown past the memory of the machine. Voxel Size, domain growth, substeps and caching decide how fast that happens.
- Voxel Size: halving it needs about eight times the memory and time, according to SideFX
- Domain: Sparse is the default; Limit Maximum Size stops the box growing out of control
- Substeps: Max Substeps with CFL controls substeps; global substeps keep every substep in memory
- Memory savers: Use 16bit Float halves field memory; cache to disk with a File Cache SOP
- Houdini 22: the old Pyro Solver DOP is deprecated in favour of the Sparse version; GPU COP Pyro is new
A Pyro sim that runs fine for 80 frames and then takes Houdini down is one of the most common problems in effects work. Early frames are cheap because the smoke is small. As it spreads, the sparse domain grows, more voxels become active, and memory climbs until the operating system ends the process.
This guide explains which Pyro Solver settings drive memory, how to cut it without losing the look, and how to cache so a crash costs you minutes rather than the whole simulation. Parameter names follow the Houdini 22 documentation.
Checked against SideFX Houdini 22 documentation for the Pyro Solver SOP, DOP Network, File Cache SOP and Performance Monitor, November 2026.
Why voxel size is the first thing to check
The Pyro Solver documentation is blunt: cutting your voxel size in half requires eight times the memory and time. The reason is simple geometry. A domain of length L at voxel size dx has (L/dx)³ voxels, so halving dx multiplies the count by 2³.
How to read the figure. Each bar is the same dense 10 metre box at a different Voxel Size. At 0.1 it holds one million voxels; at 0.05, eight million; at 0.025, sixty-four million. The orange bar is where many workstations run out of RAM. Real sims are sparse, so actual numbers are lower, but the ratio holds wherever smoke is active. That is why a sim that is fine at preview resolution can fail when you halve the voxel size for the final.
Sparse domains, padding and limits
Simulation Type > Sparse is the default. Calculations only happen where the active field says so, and the domain resizes as the smoke moves. Padding adds a margin around active areas, and Resize in Full Tiles, which is always on for sparse sims, grows the domain in 16-voxel steps.
How to read the figure. On the left is a sparse domain: only the orange tiles around the smoke are solved, so memory follows the size of the plume. On the right is a dense domain, where every voxel in the box is solved whether there is smoke in it or not. The yellow note at the bottom is the safety net: Limit Maximum Size, with a Size and Center, caps how far a sparse domain can grow. SideFX describes it as protection against running out of memory when inputs misbehave, for example a source that suddenly emits far outside the shot.
GPU acceleration has its own rule: Use OpenCL on the Pyro Solver only works when Simulation Type is Dense. Minimal OpenCL keeps the data in video memory, but it does not support simulation caching, dynamic resizing or CFL substepping, so a dense OpenCL sim can hit the GPU’s memory limit faster than a sparse CPU sim hits RAM.
Substeps: stability without wasted memory
Max Substeps and the CFL Condition work together: when Max Substeps is above 1, the CFL condition decides how many substeps each frame actually takes. There is no Min Substeps parameter. Raise Max Substeps only for fast-moving sources that tear or stream. Be careful with Global Substeps, which SideFX notes uses more memory because every substep is kept.
Settings that keep a long Pyro sim alive
The mockup gathers the settings that matter, numbered in the order to check them.
How to read the figure. Badge 1 is Voxel Size: block out at twice the final size and only go finer once timing and shape work. Badge 2 keeps the sim Sparse, and badge 3 caps the domain with Limit Maximum Size. Badge 4 keeps Max Substeps at 1 unless the CFL condition needs more. Badge 5, Use 16bit Float, halves field memory, though VDBs are only stored as 16-bit when written to disk. Badge 6 is the safety net: write the sim with a File Cache SOP, Simulation on, so each frame leaves memory and a crash only costs the frames since the last save.
Caching so a crash does not cost the whole sim
On the DOP Network you can cap the in-memory cache with Cache Memory (MB), allow it to spill with Allow Caching To Disk, and turn on Save Checkpoints so the sim resumes from the last checkpoint instead of frame one. On the File Cache SOP, the default file type is .bgeo.sc in $HIP/geo; use .vdb only if every primitive is a VDB. SideFX warns that Save to Disk in Background may use more memory, so on a tight machine save in the foreground.
Houdini 22 changes the defaults you may find in older tutorials: the Pyro Solver DOP and Smoke Solver DOP are deprecated in favour of their Sparse versions, and a new GPU-based COP Pyro workflow arrives in Copernicus (not available in Houdini Core).
Run through these before you commit to a final-resolution sim.
SaaS or IaaS for Pyro simulations
Simulations are harder to hand off than renders, because they run on one machine from frame one and must be cached before rendering.
Where to run heavy Pyro sims
A neutral comparison of the two farm models for simulation work.
| Your situation | SaaS render farm | IaaS render farm (iRender) |
|---|---|---|
| You need to simulate, not just render | Many farms only render; sim support varies | Run the sim on the server as on your workstation |
| Your sim needs more RAM than you have | Depends on the farm's node specs | Every server has 256 GB ECC RAM |
| A sim crashes halfway | Resubmit and wait | Resume from checkpoints or the last cached frame on the same server |
| Licensing | Provided for supported tasks | You bring your own SideFX licences |
Running Pyro sims on an iRender server
How to read the figure. The top row is the setup: your own PC or Mac connects by remote desktop to a whole bare-metal server with 1, 2, 4 or 8 GPUs, so you open your Houdini project there exactly as at home. The numbered row is the order of a session: register and top up, upload to iRender Drive before the server runs, start and connect, render a test frame and the job, then shut down so billing stops.
- Register an iRender account and top up; the first top-up of $50 or more gets a 100% bonus.
- Upload the .hip file and sources to iRender Drive before you start the server.
- Start a server with 256 GB ECC RAM and connect by remote desktop.
- Install Houdini once with your own SideFX licence; the server image is kept between sessions.
- Simulate with a File Cache SOP writing to disk, then render the cached volumes.
- Download the cache or frames and shut the server down so billing stops.
Frequently asked questions
1. Why does my Pyro sim crash partway through?
Usually because memory runs out as the smoke spreads and the sparse domain grows. Check the RAM graph at the crash; if it hits the top, reduce resolution, cap the domain and cache to disk.
2. How much more memory does a smaller voxel size need?
SideFX states that halving the voxel size requires eight times the memory and time, because the voxel count grows with the cube of the resolution.
3. Should I use Sparse or Dense?
Sparse is the default and solves only active areas, which saves memory. Use Dense only when you need Use OpenCL, which works only with Dense.
4. Do more substeps use more memory?
Substeps mainly cost time. Global Substeps uses more memory because every substep is kept. Leave Max Substeps at 1 unless the CFL condition needs more.
5. How do I resume a crashed Pyro sim?
Turn on Save Checkpoints on the DOP Network so the sim can resume from the last checkpoint, and write frames with a File Cache SOP so finished frames are already on disk.
6. Can I run Houdini sims on iRender?
Yes. You install Houdini on the server with your own SideFX licence and run the sim on a machine with 256 GB ECC RAM; the server image is kept between sessions.
Run your Houdini 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 caches stay installed. For Karma XPU and GPU-accelerated solvers 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).


