GPU Cache
CUDA Compute Cache
CUDA stores JIT-compiled GPU kernel binaries in the ComputeCache directory after translating PTX or other intermediate code for your specific GPU architecture and driver stack. Those cached cubin objects and related compilation artifacts speed up later launches of CUDA-accelerated apps, but they can become stale after driver updates, toolkit changes, or application updates. Kudu removes only these rebuildable CUDA JIT cache files so the next CUDA workload can regenerate fresh binaries, without touching your projects, models, datasets, accounts, or settings.
Why clean CUDA Compute Cache?
- Driver updates can invalidate cached PTX-to-binary results, causing CUDA apps to pause on first launch, recompile kernels, or briefly appear frozen while the cache rebuilds
- Stale compiled kernels from an older CUDA toolkit or app build can trigger startup errors like failed module loads or unexpected JIT recompilation loops in GPU-accelerated software
- A bloated ComputeCache consumes disk space under your profile, which users usually notice as large NVIDIA cache folders even though no new projects or data were added
- Corrupted cache entries can cause CUDA programs to fall back to recompiling kernels every run, showing repeated long initialization times instead of fast warm starts
- After switching GPUs or running the same profile on different NVIDIA hardware, cached binaries for the wrong compute capability can produce longer launches and extra compile activity before work begins
- Interrupted writes during a crash or forced shutdown can leave broken cache artifacts behind, which shows up as one specific CUDA app hanging or failing during GPU initialization while others still work
What gets cleaned
Cache paths Kudu targets
Windows
%AppData%/NVIDIA/ComputeCache |
Linux
~/.nv/ComputeCache |
Frequently asked
Common questions about CUDA Compute Cache
Free & open source
Download Kudu and reclaim your disk space.
Available on Windows, macOS, and Linux. No account required, no feature gates, no telemetry without consent. All cleaning targets are open source and community-auditable.