The post highlighting nvJPEG encoder performance is very vague as to how nvJPEG encoding is done, in fact it does not include a standalone example of JPEG encoding. In order for the data to be informative, it MUST include the quantization table used when encoding the image, as an image with a quantization table of all ones has very different performance than a standard table (for example). Would it be possible for this benchmark to be rerun using the libjpeg turbo testbench to get a more accurate and transferrable data set for future JPEG research? It includes images that are particularly taxing to encode and a clearly defined quantization table.
I tried to download the tools myself, but they won’t build and at the end of the day I don’t own a A100 (shocker), so I won’t be able to profile without the aid of Nvidia.