TMPGEnc Results
For our tests, we typically output the test files into the MPEG-4 AVC main profile, adding both video noise reduction and color correction filters. The one exception was when we output the Get Out! file in high-def without the filters, just to see what it would do to CPU utilization. You can clearly see the GPU taking a lot of the encoding load off the CPU during MPEG-2 output, but this doesn’t necessarily translate into transcoding acceleration. The 9600 GT was actually slower when rendering into MPEG-2 with GPU acceleration enabled. Curious that we see a GPU-assist improvement with the 9800 GTX but not the 9600 GT.
Back to the AVC encoding. On our 15-second file test, outputting to 1440x1080, we see CUDA giving a 19% benefit on the 9600 GT and a 45% benefit with the 9800 GTX. Here’s an interesting point. TMPGEnc also tells you the split between average CPU and GPU usage during GPU-assisted transcoding. On the 9600 GT, this was about split at 52% GPU and 48% CPU, even though our maximum observed CPU utilization in the Windows Resource Monitor was 95 percent. However, on the 9800 GTX, we saw an 11% CPU and 89% GPU split. This would seem to indicate that the application is making full use of those extra stream processors.
Just to double-check the claims about only filters being accelerated, we also ran the Get Out! clip to 1440x1080 with no filters added and saw just what you’d expect: a slight performance improvement from having GPU assistance enabled but no real speed explosion because CUDA never came into play. Without CUDA, we saw a 92% CPU, 8% GPU split on both cards.
With both filters applied, the Get Out! render had decent results. On the 9600 GT, output times dropped by five minutes with CUDA’s assistance—a 20% gain. With the 9800 GTX, we dropped eight minutes for a 32% improvement.
Honestly, I’d like to see Pegasys making better use of CUDA for encoding than just accelerating filters. Among everyday people, the former gets used a lot more than the latter. But the results here still make the case for CUDA’s potential, even if other applications do a better job of leveraging that potential.
There's a budget GeForce GPU selling in China that not even Nvidia knew it made — RTX 4010 turns out to be a modified RTX A400 workstation GPU
US to patch loopholes that allow China to buy banned AI GPUs from other countries — new regulations include national quotas on GPU exports and a global licensing system
