TensorRT RTX - Support for older RTX GPUs (Compute Capability 7.5)

[christian-wojek]

Thanks for including the TensorRT RTX library as an execution provider in 1.23.0! While the NVIDIA documentation (https://docs.nvidia.com/deeplearning/tensorrt-rtx/latest/getting-started/support-matrix.html) states that GPUs with compute capability 7.5 are supported with some restrictions the EP does only provide support for GPUs with compute capability 8.6, 8.9 and >=12.0.

• Is there a reason why 7.5 is not support?
• Is there a plan to support these GPUs in the future?

[Advait Kawale (Advait251206)]

Hi Christian,

There seems to be a distinction between TensorRT-RTX’s hardware support and the architectures currently supported by the ONNX Runtime TensorRT-RTX Execution Provider.

NVIDIA’s TensorRT-RTX documentation lists Turing GPUs (compute capability 7.5) as supported, but notes that Turing compatibility is not included in the default portable-engine configuration. Turing engines may need to be built explicitly for compute capability 7.5, and some operations have limitations—for example, FP32 GEMMs and 3D convolutions are not supported in the documented release.

However, the ONNX Runtime TensorRT-RTX EP documentation currently describes support for Ampere and later RTX GPUs. That suggests the EP’s supported architecture range is narrower than TensorRT-RTX’s overall hardware support. I haven’t found a documented explanation confirming whether this is due to an implementation limitation, validation scope, or a deliberate support decision, so I wouldn’t want to speculate about the reason.

For a Turing GPU today, the CUDA Execution Provider is a practical alternative if your goal is GPU inference through ONNX Runtime. It does not provide TensorRT-RTX engine optimization, but it can still execute supported model operations on the GPU.

References:

• [ONNX Runtime TensorRT-RTX EP: supported GPUs and usage](https://onnxruntime.ai/docs/execution-providers/TensorRTRTX-ExecutionProvider.html)
• [NVIDIA TensorRT-RTX support matrix: Turing limitations](https://docs.nvidia.com/deeplearning/tensorrt-rtx/latest/getting-started/support-matrix.html)
• [NVIDIA TensorRT-RTX engine portability and explicit compute-capability targeting](https://docs.nvidia.com/deeplearning/tensorrt-rtx/latest/inference-library/cpu-engines.html)
• [ONNX Runtime CUDA Execution Provider](https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html)

As for a plan to add compute capability 7.5 support to the ONNX Runtime EP, I’m not aware of a published commitment. It would be useful for the maintainers to clarify whether Turing support is planned and, if so, whether it requires changes in the EP or an upstream TensorRT-RTX capability.

If this answer helped or pointed you in the right direction, I'd appreciate it if you could mark it as the accepted answer so it's easier for others with the same issue to find.

Also, if you found my contribution useful, I'd appreciate it if you could check out my GitHub profile, follow me, and star any repositories you find interesting.

GitHub: https://github.com/Advait251206

[jinwoo (dev-jinwoohong)]

There is an explicit release decision in the history: PR #25611, merged July 31, 2025, disabled Turing EP devices and states that Turing would not be supported in that release. Its device filter requires compute capability major version 8 or newer.

The built-in EP still has a constructor check accepting only CC 8.6, 8.9, or 12.0 and above. The newer standalone EP is not a documented Turing workaround either: its latest release, v0.4.2, filters out devices below CC 8.0. Installing a newer TensorRT-RTX SDK alone does not remove these integration checks.

That confirms the restriction was intentional in the ORT integration. The PR does not explain the underlying engineering reason or give a date for restoring support, so it should not be read as a commitment either way.