CPU still sits at ~30% during transcription even though the GPU is confirmed in use

My machine has both an AMD iGPU and a discrete Intel Battlemage card. Whichever one I select for transcription, I can verify the GPU is actually being used, but CPU usage still climbs to around 30% for the whole job. Is that expected or is something falling back to CPU?

Resolution: That is expected. Whisper transcription cannot be fully offloaded to the GPU the way hardware video transcoding can in Plex or Jellyfin — audio decoding, chunk handling and part of the inference pipeline stay on the CPU, so there will always be meaningful CPU usage during a GPU-accelerated run. Sustained CPU at roughly a third of your cores alongside visible GPU activity is normal and is not a sign of silent CPU fallback. Only worry if the GPU shows essentially no activity while the CPU is pinned, which points at a driver or device-passthrough problem instead.