Performance of HEVC 10 bit hardware decoding.

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi. I need to decode 12 channels of HEVC UHD 50p. For this I used the nvidia quadro p5000. I ran the test and I got to decode only 3 channels on this device with 77%-83% load of decoder core (DXVA). 4 channels were decoded at a frequency below 50 fps.
In this, I have 4 questions:

  1. How much HEVC UHD 50p decoding channels I really can decode on nvidia quadro p5000?
  2. If I install 2 video card nvidia quadro p5000, will I have the ability to double the result and decode 6 channels of UHD (8 with skipping frames)?
  3. Will the result change if I use nvidia quadro p2000 instead of nvidia quadro p5000?
  4. Do nVidia has solutions that can decode more then 4 HEVC UHD 50p channels on one device?
2

note that two resolutions are defined as UHD:

[url]https://en.wikipedia.org/wiki/Ultra-high-definition_television[/url]

I assume you mean 4K UHD, i.e. 8.29MPixel

The pascal NVDEC capability is outlined here:

[url]https://developer.nvidia.com/nvidia-video-codec-sdk[/url]

So you have basically 8kx8k @ 30fps capability for NVDEC HEVC support, i.e. 64Mpixel at 30fps i.e. 1920Mpixel/s

3 of your UHD channels are using 25Mpixel @ 50fps i.e. 1250Mpixel/s 1250/1920 = 66% est. load, so your measurement of 77-83% load is in the right ballpark.

The table number /estimate of 1920Mpixel/s is an approximate number and will vary based on exact coding characteristics as well as exact clock rate of your GPU.

3

Thank you txbob. You’re right, I mean 3840x2160=8.29MPixel. Unfortunately, I still don’t understand have quadro p5000 and quadro p2000 same HW decoders performance? Both of them has same chip - pascal. Does this mean that they HW decoders have same performance?

4

I wouldn’t say p2000 and p5000 have the same chip. According to this link:

[url]Video Encode and Decode GPU Support Matrix | NVIDIA Developer

P2000 has GP106 and P5000 has GP104, however these both appear to have similar decode capabilities, and each has 1 NVDEC block per chip, so I would say they should have similar capabilities for use cases that are supported on either device. There are apparently some differences however (e.g. VP8 support), so I would suggest evaluating your use case on each device to be sure, before making any large-scale decisions.

YMMV.