Sébastien Taylor

Yeah this is measured on dataset so for this we're focused on keeping the whole pipeline busy, the VPU (JPEG), GPU, NPU, can be run concurrently and the heavy CPU portion is all in the attention and post-processing since orchestration is very low overhead. So NPU is the bottleneck here, but we're reading a dataset so we never drop frames. For a camera pipeline it depends on if you're latency or throughput focused, but also where the bottleneck lives; if it's the NPU (or other compute) then you need to decide how you want frames dropped. Could make for a great follow up topic!