How to Determine Relative Video Decoding Performance from Specs

Is there a resource for roughly determining the relative NVDEC performance of different GPUs? Does it scale similarly to the core count, or just the number of decoding units?

Specifically, I have a decoding application that performs well enough on a Quadro P6000, and the customer wants to know if they can use any other GPUs. I’m not sure how to know that without benchmarking them myself.

Hi.

We have a chart on https://developer.nvidia.com/nvidia-video-codec-sdk that shows decode performance across GPU generations. Data from this chart combined with GPU generation from https://developer.nvidia.com/video-encode-decode-gpu-support-matrix should be sufficient for you to estimate relative decoder performance. Note that # cores does not affect decoder performance since this is an independent hardware engine.

Thanks.

I really can not understand again how to compare GPUs that are not listed in https://developer.nvidia.com/nvidia-video-codec-sdk for decoding performances. Lets say T4 decoding performance is very high compared to P6000, but RTX4000 or RTX6000 is not listed, how will I know their decoding performances without benchmarking myself? Chip family? #of NVDEC chip?
Thank you.

Check again matrix it contains #NVDEC chip count (RTX4000 or RTX6000).

Performance depends:

  • stream content (used codec, resolution, bitdepth, I-P-B, BW, Q, refframes…)
  • number of streams (if you have more acceleration chips)
  • generation of acceleration chip
  • number of acceleration chips
  • frequency of acceleration chip

-> NVDEC_Application_Note.pdf

Hi,
RTX 4000 and RTX 2060 have the same chipsets TU104. RTX 4000 has 2 NVDECs, does RTX2060 also have 2 NVDECs? do they have the same decode performances?

Hi.
RTX 2060 has 1 NVDEC. For encode and decode GPU support matrix please refer https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new

In general, nvdec of same generation in any GPU will have similar decode performance. Actual performance may vary depending on clocks/memory etc.