Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU)
• DeepStream Version
I have an application that will have 100’s of rtsp cameras per zone. Currently we have a test site with 100 cameras in a single zone. Some sites will have multiple zones, say 6 zones and each zone could have 150 cameras.
My application is similar to tracking people with action recognition. Think of something like a mall to track individual people as they walk around and identify what they are doing - walking, eating etc.
I am choosing hardware and I want to make sure I have chosen the correct architecture for this.
My current approach is to use deepstream with an object detector PGIE and then pass on to a action recognition SGIE. Information will be output over kafka to allow a third-party tracker to accumulate the location information from each camera and track movement across cameras.
Inference must be done on site due to bandwidth limitations. I have chosen the A5000 GPU since it seems to give the best performance vs cost. I will need multiple A5000 since one is not enough, perhaps e.g. 10 GPUs depending on framerate.
I’m not sure whether to maximise the number of GPUs in a single server PC or have lots of smaller servers with fewer GPUs per server. I think with the deepstream / kafka approach this doesn’t matter since the third-party tracker will consolidate anyway.
Is going for A5000s here sensible? They seem to give good performance per £. There are so many cards that it’s hard to know how to pick never mind which one.
Is using multiple Deepstream instances per zone a sensible thing to do? Would one big PC be better here? One limiting factor is h264 decode, but the A5000 has two NVDEC. I think a single deepstream instance is locked to a set GPU anyway since the benefits come from the decode and inference never leaving the GPU.
Should I be using the Deepstream Triton server to “pool” GPU resources somehow instead of a deepstream process per GPU?
Should I be looking at something more enterprise-like with virtual GPUs?
Any input appreciated.