GPU Cache

CUDA Compute Cache

CUDA applications that ship PTX or other intermediate GPU code often rely on NVIDIA's driver to JIT-compile kernels for the exact GPU architecture and store the results in the CUDA Compute Cache. That cache typically contains compiled kernel binaries, device-specific code objects, and related metadata so later launches can skip recompilation. Kudu removes only these rebuildable JIT cache files from NVIDIA's ComputeCache folder, not your CUDA projects, models, datasets, settings, or account data.

Why it matters

Why clean CUDA Compute Cache?

  • Driver updates can invalidate previously JIT-compiled kernel binaries, leaving stale entries that cause a noticeably slow first run while CUDA apps fall back to recompiling kernels
  • A bloated ComputeCache consumes disk space in AppData or ~/.nv, which users usually notice as unexpectedly large per-user NVIDIA cache folders on the system drive or home partition
  • Corrupted cached code objects can trigger longer startup, repeated kernel compilation pauses, or one-off failures when a CUDA application first initializes the GPU
  • Switching to a different NVIDIA GPU architecture or using the same profile on another machine can leave behind device-specific binaries that are useless for the current hardware and only add clutter
  • After toolkit, driver, or application updates, old JIT outputs may no longer match the PTX being loaded, which shows up as extra compile time before rendering, inference, or simulation begins
  • Some CUDA workloads compile many kernels over time, so stale cache entries accumulate and make cache directories harder to manage without improving launch speed for current workloads
  • Cleaning forces the driver to regenerate fresh compiled kernels for the current driver and GPU, which often resolves odd first-run behavior without touching saved projects, training outputs, or application preferences
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.