GPU Selection Recommendations for Video Analytics

Introduction

This document provides recommendations for selecting a GPU that is fully compatible with video analytics applications. Based on these recommendations, users can make informed decisions and use the power of GPUs for their video analytics needs.

GPU Architecture

The GPU should be based on the Nvidia Pascal architecture or a newer generation.

GPU Memory Requirements

The GPU should have a minimum of 4GB RAM. While it is theoretically possible to run 2-3 analytics on 2GB, this is not recommended.

GPU RAM and CUDA Cores Calculation

A dedicated calculator will compute the required GPU memory and CUDA cores:

  • The calculated GPU RAM should be supplemented with an additional 20% capacity (no less than 1GB).
  • Similarly, the calculator will suggest the number of necessary CUDA cores. Ideally, the selected GPU should exceed the computed amount by 20% in capacity.

Decoding & Encoding Chips

The system should utilize NVIDIA GPUs which are equipped with NVDEC/NVENC decoding/encoding chips:

  • A minimum of one NVENC chip is required for video broadcasting.
  • Video stream decoding is carried out using the NVDEC chip.

Decoding Capacity of NVDEC Chip

One NVDEC chip can decode:

  • 25 video streams at h264 2MPx 25FPS
  • 50 video streams at h265 2MPx 25FPS
The resolution and FPS have a linear relationship.

Video Stream Calculation

The following formula determines the number of video streams decoded by one decoding chip for h264: 25 video streams * 25 FPS * 2MPx = 1250 units.

1250 units / your resolution in MPx (e.g. 4MPx) / your FPS (e.g. 15 FPS) = the number of video streams that can be decoded (20 video streams for this example).

  • Similarly, for h265, the calculations follow the same formula: 50 video streams * 25 FPS * 2MPx = 2500 units.
  • When the system has two NVDEC chips, it can decode twice the number of video streams. This relationship is also linear.

GPU Architecture and Chip Count Verification

The number of NVDEC/NVENC chips and the supported architecture can be verified via this link: https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new.