I was trying the new ML Train GSplats node in Houdini 22 on macOS, but it immediately fails with:

ERROR: No CUDA GPUs detected. GSplat training requires a CUDA-capable GPU.

So once again, a new Houdini feature appears to be Nvidia only and therefore unavailable on macOS.

Did I overlook an alternative backend or setting for Apple Silicon, such as Metal or MPS, or is GSplat training currently CUDA only?

Maybe someone from SideFX or anyone who has tested this on a Mac can provide some insight.

chf
I was trying the new ML Train GSplats node in Houdini 22 on macOS, but it immediately fails with:

ERROR: No CUDA GPUs detected. GSplat training requires a CUDA-capable GPU.

So once again, a new Houdini feature appears to be Nvidia only and therefore unavailable on macOS.

Did I overlook an alternative backend or setting for Apple Silicon, such as Metal or MPS, or is GSplat training currently CUDA only?

Maybe someone from SideFX or anyone who has tested this on a Mac can provide some insight.

Hi,

I came to ask the same thing - after watching an entagma video about gpslats I thought I'd try it.

I dug around the python files and did find this in one of the multiple train.py files that is created.

# Default to 'cuda' if it's available, or mps, else use CPU
if not device_name:
if torch.cuda.is_available():
device_name = 'cuda'
elif torch.backends.mps.is_available():
device_name = 'mps'
else:
device_name = 'cpu'

I even edited the contents of the ml_traingsplats and changed the path of the external python script file to one that had this line it but got a different error -

No module named 'torch'

So it looks like it should work without cuda somehow.

That does sound like an oversight on SideFX’s side. The training code itself clearly seems prepared to fall back to MPS on Apple Silicon, but Houdini appears to stop the process earlier with its CUDA check.

I remember using a few GitHub projects that were primarily built around CUDA, but after installing the correct PyTorch build for Apple Silicon and a couple of additional dependencies, they worked through MPS. I just have not had the time to dig through Houdini’s generated environment and try the same thing here.

I would also have expected SideFX to handle that setup automatically on macOS, especially if the underlying script already contains an MPS fallback. Hopefully it is just a missing dependency or an overly strict validation check rather than the actual training implementation being CUDA only.

EDIT: I sent an email detailing the issue to support@sidefx.com

Got an answer from SideFX support:

"I can confirm that ML Train GSplats requires a CUDA capable GPU (of cuda capability >= 7.0 & <= 12.0) as the error message points out. It is not intended to work on stock macOS."

Kind of disappointing to see this locked to proprietary tech, unless I'm missing something, I don't think there's a hard technical reason for it. A lot of the underlying stack (PyTorch, etc.) has non-CUDA backends available, as SteveWatsonHoudini's find above shows. I get that porting something like a full OptiX-based pipeline to Metal/MPS would be significant work, but this doesn't strike me as being anywhere near that level of effort, happy to be corrected if there's a technical reason I'm not seeing.

GSplats is one of the headline features of H22, so it's a bit of a letdown for Mac users. At the very least, it would be good if the marketing/docs made it explicit upfront that this feature is CUDA-only, so people know what they're getting before they try it. Showcasing a flagship new feature that simply doesn't run for a non-trivial share of the user base, without a clear disclaimer, feels a little misleading.

chf
Got an answer from SideFX support:

"I can confirm that ML Train GSplats requires a CUDA capable GPU (of cuda capability >= 7.0 & <= 12.0) as the error message points out. It is not intended to work on stock macOS."

Kind of disappointing to see this locked to proprietary tech, unless I'm missing something, I don't think there's a hard technical reason for it. A lot of the underlying stack (PyTorch, etc.) has non-CUDA backends available, as SteveWatsonHoudini's find above shows. I get that porting something like a full OptiX-based pipeline to Metal/MPS would be significant work, but this doesn't strike me as being anywhere near that level of effort, happy to be corrected if there's a technical reason I'm not seeing.

GSplats is one of the headline features of H22, so it's a bit of a letdown for Mac users. At the very least, it would be good if the marketing/docs made it explicit upfront that this feature is CUDA-only, so people know what they're getting before they try it. Showcasing a flagship new feature that simply doesn't run for a non-trivial share of the user base, without a clear disclaimer, feels a little misleading.

Hi,

that's disappointing, I know splats can be made on Mac OS because Ive been using an app called 3D splat app - which includes colmap as part of it's tool. Just would be nice to compare it to houdini to see if it's better/different/worse.