Can I Run Yolo v3 without GPU in a server?

[Albert2020666]

Hi, I am trying to implement Yolo v3 on an MS server without GPU, but I'm not sure if it is sufficient to run the system for detection. Does the detection speed depend on how many targets for detection? For example, if I only detect one object, is it faster than detecting multiple objects?

Thank you so much

[jklemmack]

I actually do this, leveraging the CPUs. Not an issue. It will use all cores on the CPU, and the processing time is constant based on the configuration model, not the complexity of the image.

Rough numbers, an Intel i7 9700K takes about ~1-2 seconds to process my 608x608 network.

[Albert2020666]

I actually do this, leveraging the CPUs. Not an issue. It will use all cores on the CPU, and the processing time is constant based on the configuration model, not the complexity of the image.

Rough numbers, an Intel i7 9700K takes about ~1-2 seconds to process my 608x608 network.

Thank you for your reply. So, it is actually doable. My system has around 10 ip cameras for input and detects one specific object. Do you think the system is able to run without GPU? Thank you.

[jklemmack]

In my experience the number of objects outputted matters less than the input size of the network. Sticking with that 608x608 example, then you'd only have 1 frame per second across 10 cameras. Rolling through the cameras, each one could detect 1 image per 10 seconds. If that doesn't fly, then either get more CPUs/cores, change the network size (maybe Yolo Tiny?), or pony up for any type of CUDA-supported graphics card. Even low end devices are going to give you a big boost.

On the far end, we are also using a 1024x576 network at 25 FPS using a Titan RTX.