Hello,
we’re working with customer to find proper hw specification.
Customer using HP server with Xeon 6760P CPU and RTX6000 GPU.
he controll IP camera over 120ea.
Now we working to replace this server with Hailo-8.
we have 104 TOPs card(Hailo8 x 4)…
Can you recommend the target CPU & number of NPUs for this project?
we have plan to use each Hailo chipset for several IP cam.
I beive Hailo team have semilar experienes.
I’ll wait your advise.,
thanks
It’s hard to tell without the full details on the application, video resolution, analytics FPS, analytics pipeline, and so on.
Under a lightweight application assumption that the only AI model is a small detector like yolov8n running at 10 FPS per stream, each Hailo-8 runs at >1000FPS, so it can even handle 100 streams. Choosing a bigger detector such as yolov5m will result in 242/10=24 streams per Hailo device. You can use the Model Explorer to see the FPS of each AI model on Hailo-8.
Running a more complex multi-stage AI pipeline will probably lead to 8-16 streams per Hailo-8 device, but again it heavily depends on the assumptions. For example, using tiling to detect smaller objects will again raise the required TOPS.