GPU processing

On Business or Enterprise plans, you can ask any embedding or detection endpoint to run on a GPU (a graphics processor that can speed up the work). The output format, the watermark and the output check are the same as on CPU.

Default

CPU

GPU cost

3× normal credits

If GPU is not ready

CPU at normal cost

Request GPU processing

Add accelerator=gpu to a sync or async request. Use accelerator=cpu to request CPU explicitly.

curl --fail-with-body "https://api.etchv.com/watermarks/videos/async?accelerator=gpu" \
  -H "X-API-Key: $ETCHV_API_KEY" \
  -H "Idempotency-Key: launch-video-partner-42" \
  -F 'file=@launch.mp4' \
  -F 'data={"asset":"launch-video","recipient":"partner-42"}'

The dashboard, MCP and automation integrations offer the same choice. How much faster a GPU is depends on the file, because reading and writing the file still run on CPU. See performance benchmarks.

Credits and fallback

Etchv holds 3× the normal credits when the request starts. If the work runs on CPU instead, Etchv charges the normal amount and releases the rest. If the operation fails, all held credits are released.

GPU readiness

An active Business or Enterprise workspace keeps GPU capacity running and ready (“warm”). Extra capacity can take a few minutes to start. Requests run on CPU while it starts.

Check readiness

curl --fail-with-body "https://api.etchv.com/accelerators/gpu" \
  -H "X-API-Key: $ETCHV_API_KEY"

Checking does not start a GPU. Use a key with watermarks:embed or watermarks:detect.

Start the GPU

curl --fail-with-body -X POST "https://api.etchv.com/accelerators/gpu/wake" \
  -H "X-API-Key: $ETCHV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"webhook_id":"wh_0123456789abcdef0123456789abcdef"}'

Waking costs no credits and needs a Business or Enterprise plan. Once ready, the GPU stays warm for 15 minutes. GPU work in progress keeps it running longer. Waking it again extends the 15 minutes.

If you pass the optional webhook_id, that endpoint receives accelerator.ready. It is sent right away if the GPU is already ready. Without a webhook, poll the readiness endpoint. Both endpoints return accelerator, status, warm_until and ready_at.

See which accelerator ran

The Usage dashboard also shows the accelerator.

Errors

GPU requests share your plan’s rate limits.