AMD Instinct MI300X 192GB
The AMD Instinct MI300X offers 192 GB of HBM3 per card. It can appeal to a creator preparing a generative workflow under ROCm who needs large capacity on a single GPU. Each RenderParcel lot contains one MI300X.
- GPUs per batch
- 1
- Memory per card
- 192 Go HBM3
- Stock
- 8 units
Crypto payment without KYC · your tools, your creation.
Set up my rental 8 units in stockStart from an AMD-compatible workflow
AMD documents a ComfyUI and PyTorch combination on MI300X. That provides a technical starting point, without meaning that every community node and every recipe designed for CUDA will work as is. Identify the extensions that use NVIDIA-specific components before moving your project.
Build a minimal test with the model and the essential operations. Then add the conditioning and output processing. This progression lets you isolate an incompatibility instead of confusing it with a memory issue.
192 GB for a complete recipe
The capacity can accommodate projects that exceed the 80 or 141 GB envelopes, provided the software and the model support that configuration. Assess the needs of the entire pipeline: model weights, intermediate data, dimensions, and output steps.
Keep the numeric precision and the settings in your notebook. If you compare two environments, also observe how the images look; do not assume identical results from the same seed and the same model name.
A clear boundary with Blender/Cycles
The MI300X uses the CDNA 3 architecture. The published HIP support for Cycles targets the AMD RDNA families; it does not allow advertising the MI300X as a compatible Blender/Cycles choice. For a scene to render with Cycles, look instead at an RTX 6000 Ada or another appropriate graphics card.
For ComfyUI with CUDA dependencies that are hard to replace, the 141 GB H200 may be an alternative to consider. The choice should rest on recipe compatibility as much as on the advertised capacity.
Check the nodes before moving production
Draw up a short list of the essential components: model, custom nodes, GPU operations, and Python packages. For each one, check the ROCm path; then test the full graph and keep the environment that succeeds. Choose 3, 7, or 30 days based on that measurement, the number of tests, and the adaptation work still needed.
The quote must match the MI300X and the number of cards you want. After the order details, the no-KYC crypto payment takes place in the same flow. Keep an export of the environment along with the retained images: it will let you tell a recipe change apart from a dependency change when you pick the work back up.