Copy link to clipboard
Copied
...on my main Intel i9-14900K system.
The specs are:
The integrated Intel UHD Graphics 770 is enabled on this system.
I turned on and off hardware decoding and hardware encoding, and I also set the hardware decoder to Intel (Quick Sync) only or Nvidia (NVDEC) only. The hardware encoding is locked to Nvidia (NVENC) only or software only (Intel Quick Sync hardware encoding is not possible with a discrete non-Intel GPU installed).
The overall results, in order from slowest to fastest (based on my PugetBench for Premiere Pro scores using the Standard preset):
The result with Premiere Pro set to the default decoder and encoder settings was equal to that achieved with the Nvidia decoder enabled and hardware encoding enabled due to the default decoder setting, which was supposed to utilize both the Intel and Nvidia decoders, was broken in Premiere Pro 25.1.
Please note that my findings are for my particular system only. If your system has a weaker CPU, then the hardware decoding performance will always beat software-only decoding. It just so happens that my i9-14900K is more than powerful enough to handle software decoding.
If you have the chance to play around with the decoding and encoding settings, please share your findings in this discussion.
Randall
Copy link to clipboard
Copied
So if you've an i9-49000k, the graphics card is not that important for rendering?
Copy link to clipboard
Copied
The GPU is still important in rendering. In all of my tests the discrete GPU is utilized in rendering where it can, and all of my testing had the GPU set to hardware encode H.264 and HEVC by default. The lower score for decoding (called "Processing" in PugetBench) with the Nvidia hardware decoding enabled versus software-only decoding is likely due to the inherent bottlenecks in the PCIe interface (as well as all other full-duplex interfaces) – in testing, simultaneous transfers through a full 16-lane PCIe bus in both directions create some latencies within, particularly when each of the simultaneous transfers use the entire lane width of the interface.