very clever and thanks for explaining. for gpu-bound processes, which are common ffmpeg use cases, this is a great approach