Upgrading to GTX 970 not much faster, is this normal?

general questions about Neat Video
Post Reply
Trixter
Posts: 15
Joined: Wed Jun 06, 2012 2:34 pm

Upgrading to GTX 970 not much faster, is this normal?

Post by Trixter »

I mainly restore 12-bit captures of 480i material, so lots of 720x480i 32-bit work with neatvideo. I've used Premiere CS5 with a GTX 470 for years with these kinds of performance times:

720x480i 32-bits radius 1 75% of available GPU RAM: GPU only = 22.7fps, CPU + GPU = 26.3fps

I upgraded the video card to the GTX 970 which over has 3 times the cores and memory bandwidth, but when I test, I see only a tiny improvement:

GPU only = 30.3fps, CPU + GPU = 31.3

Why such a small improvement? Any ideas? This was very disappointing to see after spending nearly $400 on a new video card.

(All CPU+GPU times were done by choosing Optimize and accepting the fastest combination. Also, I know the last 0.5G of the 970's RAM is slower than the rest so I was careful to tell it to use only the first 75% of GPU RAM.)
NVTeam
Posts: 2745
Joined: Thu Sep 01, 2005 4:12 pm
Contact:

Post by NVTeam »

> Why such a small improvement?

I think the main reason is the small frame size. When the amount of data to be processed is small, the GPU processing overhead (which is always present) becomes non-negligible in comparison with the actual frame processing done by GPU. Even if the frame processing is much faster, the overhead is still the same.

For comparison I recommend to check the speedup (GTX 470 vs GTX 970) on a larger frame, at least 1920x1080 or even larger. Then you will see the actual difference between GPUs.

Vlad
Trixter
Posts: 15
Joined: Wed Jun 06, 2012 2:34 pm

Post by Trixter »

I have confirmed that larger frame sizes show better results.

Unfortunately I still seem to be lagging behind -- the GTX 970 is one of the best cards you can get right now, yet I'm posting speeds under that of a 780 Ti. (I submitted results at http://fifonik.com/nv/ for verification).

Maybe my CPU is too old/slow, which contributes to the processing overhead...
Post Reply