Older AMD drivers on v4? (also a speed question)

resolve technical issues related to use of Neat Video
Post Reply
Zach
Posts: 38
Joined: Sat Jun 01, 2013 12:37 pm

Older AMD drivers on v4? (also a speed question)

Post by Zach »

I was curious about something and wanted to do a test.

On my other machine, where I have older "pre-bug" AMD drivers, I wanted to see if v4 would run faster than on my current rig, as its been mentioned that AMD GPU processing may run slower due to these bugs.

However with those drivers I can't use v4, as they fail an openCL version check apparently. The Catalyst version I have on this other machine is 13.1 I kept it around and haven't upgraded that machine because of the NV3 issues mainly.

Was this intended ?


I really wanted to see what kind of results I would get trying v4 on the older drivers (that worked fine with unpatched Nv3) as v4 runs pretty slow for me on my main machine with recent beta drivers, although I do like the results I am getting from it.

Part of what prompted this was in v4 even doing a benchmark I tend to get about half the reported speed during an actual render. i.e benchmark would report around 10fps or so, but I might max out at around 5fps or so (and sometimes drop just below 2fps) with 4 cores + GPU enabled. I don't think I've ever seen actual rendering speed hit the benchmark value.
NVTeam
Posts: 2745
Joined: Thu Sep 01, 2005 4:12 pm
Contact:

Re: Older AMD drivers on v4? (also a speed question)

Post by NVTeam »

Zach wrote:The Catalyst version I have on this other machine is 13.1
That may be too old (and that may be causing the check to fail). The recommended version of Catalyst is Catalyst 14.9, it doesn't yet have those new bugs added in 14.12.

So I recommend to install 14.9 on both machines and run the tests with that version.
Zach wrote:Part of what prompted this was in v4 even doing a benchmark I tend to get about half the reported speed during an actual render. i.e benchmark would report around 10fps or so, but I might max out at around 5fps or so (and sometimes drop just below 2fps) with 4 cores + GPU enabled. I don't think I've ever seen actual rendering speed hit the benchmark value.
It is normal that the speed measured by Optimize is different than the speed of render in the host application. The reason is simple: Optimize measures the speed of Neat Video alone, while the render in the host application involves much more than Neat Video alone. The render process run by the host application typically involves: decoding the original clip, any pre-processing required by the host application (like color space conversions and the like), sending data to the filter plug-in and back, any other filters that may be used, Neat Video itself, any required post-processing, final encoding and saving the output file. Therefore that multi-step process is always slower than any of its individual steps.

Vlad
Zach
Posts: 38
Joined: Sat Jun 01, 2013 12:37 pm

Post by Zach »

Thanks I will look into getting that version of Catalyst.

I will probably want to set up some kind of repeatable test with a small file, in general is it safe to do this by cutting a small section out of a full video I've already made a DNP/NFP set for, or would I need to re-profile the small clip?
NVTeam
Posts: 2745
Joined: Thu Sep 01, 2005 4:12 pm
Contact:

Post by NVTeam »

It is just fine to cut out a section. Or just set up a new noise profile. I believe the accuracy of the profile is not critical for this test.

Vlad
Zach
Posts: 38
Joined: Sat Jun 01, 2013 12:37 pm

Post by Zach »

I did some limited testing and I do appear to gain a little speed from reverting to the 14.9 drivers.

So I went back and ran a full render without making any changes and I gained somewhere around 1 - 1.5 hour reduction on a render time for that particular video vs what it took me on the newer drivers previously.

I'm currently running a render right now and its humming along at around 10fps or so with an estimated 3.5 hour total render time although that could go up or down. The movie is roughly the same length as the previous one I tested for the speed gain.

I guess its worth noting, on this current render I gained some more speed by reducing temporal range to 1.



I know the temporal filter is definitely the biggest contributer to slow performance in what I am doing, and I'm curious if its possible to get more tweakable options for it in the future (quality / speed trade off). Maybe the option to use less intensive algorithms or something? Or open up tweaking of values that may be hidden?
NVTeam
Posts: 2745
Joined: Thu Sep 01, 2005 4:12 pm
Contact:

Post by NVTeam »

The most important parameter of the temporal filter affecting the processing speed is the radius of course. It is already available for tweaking. Any other parameters are secondary and not likely to help in achieving any better speed without sacrificing the quality. It may be possible to use some simpler filters instead, but that is something to consider for future updates.

Thank you,
Vlad
Post Reply