I was looking for Deepstream benchmarks for V100 but wasn’t able to find it. I want to process 100+ camera feeds from Deepstream and I wanna know How many streams can a Single GPU can process? The benchmarks for T4 States that it can process ~35 streams together but utilizes only 35% of GPU power (https://docs.nvidia.com/metropolis/deepstream/4.0.2/dev-guide/index.html#page/DeepStream Development Guide/deepstream_performance.html).
So my main question is can I run multiple docker containers and process 70 Input camera feeds from T4?
or
Does it utilizes only 35% GPU and cannot take more load?
Hi @lakshay.chhabra,
Two main factors could affect the max supported streams,
the first is the max decoding perf, you can find T4 perf rrom NVIDIA VIDEO CODEC SDK | NVIDIA Developer
the second is the inference perf, this varies based on what model(s) you will run. Using TensorRT to run the model can get the best perf, you can use TensorRT to profile the perf - max fps that you model(s) can run.
If your pipeline includes other components, e.g. encoding, you also need check if its perf can support your encoding requirement.