CUDA-conan roadmap

[jcar87]

🗺️ cuda-conan Roadmap

This issue tracks planned work for the cuda-conan repository. Check items will be updated as work progresses — comments and feedback welcome, in particular to help prioritize.

---

CUDA Toolkit versions

• Add cuda-toolkit 13.3.x for Windows (x86_64) and Linux (x86_64, arm-sbsa)
• Add cuda-toolkit 13.4 developer preview for Windows (x86_64, arm64)
• Add cuda-toolkit 13.4 developer preview for Linux - waiting on tarball installers (x86_64, arm-sbsa)

CUDA Toolkit recipe features

• Add cross-build support to the 12.9 toolkit (currently implemented for 13.x only)
• Add option in the 12.x versions to distinguish between arm-sbsa and aarch64-jetson targets
• Ability to configure exposing the cuda-compat driver to LD_LIBRARY_PATH when needed (configurable by user first, but could be automatically determined)
• Implement the validate() method in the cuda-toolkit recipe to warn early about incompatible compiler versions - consider Conan config variable such that the restriction can be bypassed along with passing -allow-unsupported-compiler
• Add attribute build_policy = "missing" such that the recipes automatically generate the package without the need for a explicit --build=missing flag

Other CUDA-X libraries

• add Windows arm64 variants once released

---

Tutorials & integration examples

• Cross-building for Linux arm64 — sbsa (13.x) and aarch64/sbsa (12.x)
• Cross-building for Windows arm64 using the 13.4 developer preview
• Configuring target CUDA architectures via CMAKE_CUDA_ARCHITECTURES
• Building a library against a matrix of CUDA versions (e.g. libtorch × CUDA 12.x and 13.x)

---

Consumer recipe variants (CUDA-enabled)

• a recipe for cuda-samples that uses CUDA Toolkit from Conan
• whisper-cpp
• zfp
• opensubdiv
• sleef
• opencv

---

Repository maintenance and CI

• test_package: work around lack of libcuda.so when building on CI hosts without a GPU driver (we could expose the stub, if only to test the test_package launches)
• add CI around all recipes as well as a build matrix that includes Linux+arm64, cross-building, building the same recipe against multiple CUDA versions

Integration with Conan client (medium term)

• consider whether cuda setting (with cuda.version, cuda.arch) is the right approach
• profile plugin for developers with a GPU such that a profile can set the max supported CUDA version (as per the driver), as well as the required GPU arch
• compatibility plugin: given a requested GPU architecture, locate a binary that is compatible (prioritise the exact architecture, if not, compatible PTX)

Based on user demand

• Versions of CUDA Toolkit older than 12.9
• Older versions of the libraries (e.g. cuDNN)
• Explore ability to bypass recipe with [platform_requires] to fall back on Toolkit and libraries installed at the system level

---

Last updated: July 2026. To suggest additions, leave a comment below.