Spark Desktop Performance Caches
Spark Desktop stores several Chromium and Electron performance caches under its local app data folder to speed up startup and rendering. These include V8 compiled code in Code Cache, GPU process caches in GPUCache, and DawnGraphiteCache and DawnWebGPUCache entries for WebGPU and modern graphics pipeline state; after app updates, Chromium changes, or graphics driver changes, those files can become stale and stop matching the current runtime. Kudu removes only these rebuildable performance caches so Spark can regenerate fresh ones on the next launch, while preserving conversations, account data, settings, and saved work.
Why clean Spark Desktop Performance Caches?
- Invalidated V8 compiled code after a Spark Desktop or Electron update can make the app pause or feel unusually slow on first open until fresh bytecode is generated
- Stale GPUCache entries from a previous graphics driver version can cause black panes, flickering UI, or delayed window redraws when Spark uses hardware acceleration
- Outdated DawnWebGPUCache data can break newer rendering paths and show up as blank sections, visual glitches, or unstable animations in the interface
- Old DawnGraphiteCache pipeline state blobs can force repeated shader recompilation, which users notice as stutter when opening views or switching between heavy screens
- Corrupted cache files in Spark's Chromium runtime can trigger repeated GPU process crashes, leaving the window white or unresponsive until the cache is rebuilt
- Oversized performance caches waste space in %LocalAppData% and can keep Spark carrying around obsolete compiled artifacts long after updates, with no benefit to current builds
Cache paths Kudu targets
Windows
%LocalAppData%/Spark Desktop/Code Cache |
%LocalAppData%/Spark Desktop/GPUCache |
%LocalAppData%/Spark Desktop/DawnGraphiteCache |
%LocalAppData%/Spark Desktop/DawnWebGPUCache |
Common questions about Spark Desktop Performance Caches
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.