Video card (Tesla V100 OR Tesla T4) for the real-time video stream indexing server

Hi! Help me pleas choose a video card (Tesla V100 OR Tesla T4) for the real-time video stream indexing server, which processes the video sequence broadcast by surveillance cameras, detects images of people’s faces in it and extracts feature vectors from them - biometric templates suitable for further manipulations: monitoring, identification (search) and verification.

There is the catch that none of these GPUs are actually capable of doing the indexing even remotely in “real-time” at full frame rate and resolution. Marketing numbers published by NVidia and partners usually refer to down-sampled images in QCIF resolution and alike, but that has no relevance for real life applications where the objects are not already screen-filling to begin with but you actually require full stream resolution.

If your streams are e.g. mostly 4k, then a realistic expectation would be that out of ~400 decoded frames per second (decoding performance of T4), you can afford to analyze only 5-20, depending on your expectations about accuracy.

Things only start looking nicer if you limit yourself arbitrarily to access control systems and alike, where down-sampling before analysis becomes a viable choice.

Good day! Thank you so much! Please recommend a graphic card capable of recognizing faces and images in real time.