Stay tuned for the final CUDA 13 release expected at SC26 (November 2026), which promises to redefine GPU programming once again. In the meantime, test CUDA 12.9 in your development pipelines, but keep a production environment pinned to 12.8 until your entire dependency stack validates the update. For official release notes, visit NVIDIA CUDA Toolkit Downloads .
cuda-toolkit-upgrade check --project-dir ./my_cuda_code --target-version 12.9 The easiest upgrade path is via NVIDIA’s official Docker images: cuda toolkit update news
| Workload | GPU Used | Performance Improvement | | --- | --- | --- | | Llama 3 70B Inference (FP8) | H100 | +12% (due to better Transformer Engine) | | cuQuantum Circuit Simulation | H200 | +22% (new sparse state-vector ops) | | AMBER Molecular Dynamics | A100 | +8% (faster atomistic kernels) | | Stable Diffusion 3.5 (bfloat16) | RTX 5090 | +15% (optimized attention kernels) | Updating CUDA toolkits is not trivial. Here’s a safe pathway: 1. Use NVIDIA’s New cuda-toolkit-upgrade Tool NVIDIA released a utility in March 2026 that scans your project’s CMake/Makefiles, identifies deprecated APIs, and suggests replacements. Stay tuned for the final CUDA 13 release