doing a benchmark to compare performance between hailo and other devices. for FPS, is it more accurate to use batch size 1, or 8?
Hi @Rowan_Chin,
I’d recommend the following:
- Use batch size 1 - Hailo’s official FPS numbers are measured at batch size 1 with single-context HEFs compiled in performance mode, and typically this is the common figure.
- Make sure your HEF is from the Hailo Model Zoo - these are compiled with performance optimization enabled.
- If the other devices also support batched inference, consider reporting both batch 1 and batch 8 - so you can see latency-optimized vs. throughput-optimized scenarios. Note that on Hailo,
--batch-sizecontrols how many frames are processed before a context switch - it’s not traditional batch inference like on a GPU. You can check if your model uses multiple contexts withhailortcli parse-hef your_model.hef - Use
hailortcli benchmark your_model.heffor the easiest way to get both HW-only and full-system FPS in a single run.
Thanks,