GPU Cache
CUDA Compute Cache
CUDA stores a per-user ComputeCache of JIT-compiled GPU kernel binaries so applications using PTX can skip recompiling device code on every launch. Those files are generated when a CUDA app, library, or framework translates PTX into architecture-specific machine code for your current GPU and driver, and they can become outdated after driver changes, toolkit updates, or app upgrades. Kudu removes the cached CUDA binary blobs from NVIDIA's ComputeCache directory, forcing fresh compilation without touching your projects, models, datasets, accounts, or settings.
Why clean CUDA Compute Cache?
- Driver updates can invalidate cached JIT binaries, leading to long startup pauses or repeated recompilation messages the next time a CUDA app initializes the GPU
- Stale PTX-to-binary cache entries from an older toolkit or library version can cause CUDA programs to fail at launch with module load or invalid device code errors
- Corrupted cache files in ComputeCache can trigger crashes or hangs during the first GPU workload, especially when an app is loading kernels before a render, training run, or simulation starts
- Large accumulated kernel caches from multiple frameworks and GPU architectures waste disk space in the user profile and can noticeably grow after testing many CUDA builds
- Removing obsolete cached binaries helps when one CUDA application works on one machine but another shows unexplained first-run stalls because it keeps reusing bad JIT output
- After changing GPUs or moving between systems with different compute capabilities, old cached binaries can cause extra fallback compilation attempts and visible delays before workloads begin
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.