Need Hardware Detail for processing of 50 cameras with yolov3-tiny.

[zhaochunguang1993]

Hey, @AlexeyAB, I need your helpness.

I trained the models for ALPR, it consists of 2 yolov3-tiny models. With these, I have tried with 10 IP cameras (rtsp) in gtx1060 GPU, core i5-7400(4 Cores, 3GHz).

For video decoding, i compiled the opencv4.2 with Nvidia Video Code, so frame decoding is processed in GPU, cpu usage is decreased.

Processing is follow as below. (i developed using Python language)

• pre-processing + videocapture (one thread per one camera)
• network inference (2-4 darknet instances )
• post-processing + show results (one thread)

After that, I got 5 FPS in this test.

In your opinion, is this correct result?, What should i do in order to get more fps in this hardware? If i plan to process the 50 cameras with 5-10 FPS and plan to use 2 GPUS, then which cpu and gpu can you recommend to me? . I will wait your reply. Thanks.

[AlexeyAB]

I would recommend you to use 2 x RTX 2080 Ti.

And CPU AMD Ryzen Threadripper 2970WX - 24 Cores

---

But may be will be enough CPU AMD Ryzen Threadripper 2920X - 12 Cores it depends on next questions:

• What GPU-usage do you see in GPU-Z utility?
• What CPU-usage do you see currently?

---

• gtx1060 - 4 TFlops-SP
• rtx 2080Ti - 13 TFlops-SP (100 TFlops-TC)

So 2 x rtx2080Ti is about ~8x faster than 1 x gtx1060.

2 x RTX2080Ti can process 50 cameras with 8 FPS.

[zhaochunguang1993]

Thanks for your reply. In my test, CPU is about 100% (decoding is processed in gpu), GPU is about 40%. I don't know why GPU usage is low. In your opinion, how much fps can i get more in this test ?

[AlexeyAB]

In my test, CPU is about 100% (decoding is processed in gpu), GPU is about 40%. I don't know why GPU usage is low.

Because CPU is a hardware bottleneck. Try to use better CPU.

---

In your opinion, how much fps can i get more in this test ?

2 x RTX2080Ti + CPU AMD Ryzen Threadripper 2970WX - 24 Cores can process 50 cameras with 8 FPS.