Using 8MP video comes with trade-offs.  You double your pixel resolution and thus facial recognition performance will produce significantly better results.  But this will come at a cost of performance as explained below.


Some info about Tesla and A-series generation GPUs and high camera resolution:

  • Nvidia A-series cards do not come with drivers that perform decode on the GPU.  So in this case Decode must be performed on the CPU.  This takes away valuable CPU resources that are needed by SAFR
  • The amount of processing needed for decode varies by the resolution of the video.  At 4MP and below decode processing is approximately same as CUDA processing requirement (SAFR uses CUDA for Facial recognition).  But at higher camera resolutions, Decode performance increases significantly.  See screenshot below to demonstrate this.  This means:
    • If processing on CPU, you will take significantly more % of CPU time to decode input video
    • If processing on GPU, video decode will become the bottleneck for processing when going to 8MP video and you will get less overall cameras on a single GPU.


A screenshot of a computer

Description automatically generated




You can license Nvidia Virtual GPU licenses in order to obtain drivers that perform decode on the GPU.   See this articlefor more information.


Even if you do obtain the Nvidia Virtual GPU licesnes, using 8MP video feeds will still limit your overall capacity per GPU.