faster cuda card?

[sharkbite]

Folks this is sort of a h/w question but related only to video conversion.

I have a Phenom 1090t mated to a gts250 that I use mainly to convert captured TV
Shows for viewing on a tablet.

I use either media coder or free make because they support CUDA.

I get conversion speeds of about 7.5x But neither the CPU or gpu is particularly stressed.

I am thinking of upgrading to a gtx650 cause they are real cheap and have a heap more cores than the 250.

Just wondering if anyone has done a similar upgrade and have you noticed a decent increase in speed?

Cheers,

[deadrats]

Folks this is sort of a h/w question but related only to video conversion.

I have a Phenom 1090t mated to a gts250 that I use mainly to convert captured TV
Shows for viewing on a tablet.

I use either media coder or free make because they support CUDA.

I get conversion speeds of about 7.5x But neither the CPU or gpu is particularly stressed.

I am thinking of upgrading to a gtx650 cause they are real cheap and have a heap more cores than the 250.

Just wondering if anyone has done a similar upgrade and have you noticed a decent increase in speed?

Cheers,

the gtx650 uses the kepler gpu, prior to this gpu all video encoding on nvidia cards was done via the cuda cores, thus core count mattered,one of the problems was that on a single gpu setup encoders that really maxed out the cuda cores wouldbring the system to its knees as the desktop is accelerated by the gpu, not to mention that power consumption was through the roof. as of kepler nvidia included a hardware based h264 encoder called nvenc, not to be confused with the nvenc h264 encoder that ran on the cuda cores.

you can read about it here, all the way at the bottom:

http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf

bottom line is that a kepler based gpu is the way to go if you wish to use an nvidia accelerated encoder, so yes, pick up the gtx650, you will be glad you did.

[sharkbite]

Thanks dead rats....

Will give it a go .....


Would also dearly love to know how to stress the current setup,
As both media coder and free make only use about half the available CPU/gpu
(Well according to task manager and gpu-z anyway)

I can use up more by doing parallel encodes, but this is only useful
If you have more than one file to encode!

It would be nice if there was some s/w that used all the available resources to give
Me 15x ....

[deadrats]

you don't want any software using 100% of the resources available, that would bring your system to its knees and you wouldn't be able to do anything else with it.

what you could do with media coder is use the mencoder decoder, go into the advanced settings and manually adjust the number of threads used.

[sharkbite]

Thanks matey....

Since the machine is not used for anything else i'm not really fussed about concurrent usage. Would much rather it took the available resources to work faster

Also the default decoder for iPad profile is mencoder,and it uses all cores as default.

Setting the cuda output to "full offload' does not do much either....

Cheers,

[sharkbite]

Well I bought a galaxy gtx650 OC.....

Its not any faster......

In fact if anything its a tad slower.......

Gpuz reckons it now uses 70% of the gpu when doing a cuda encode.

Forcing media coder to use nvenc results in taking twice as long......

I s'pose at least the new card is quieter and runs cooler....using less power.....

But as I write this....a faster cuda card is not improving the speed of encoding

[deadrats]

there's something wrong with your configuration, everything i have read on the kepler gpu's with the new hardware h264 encoder says that it smokes previous cuda based h264 encoders. which drivers are you using, the official nvidia reference drivers or the one's from the card manufacturer?

[sharkbite]

Matey.....
I'm using 314.07 nvidia reference drivers.
I'm using the default settings for iPad as supplied by media coder......
Except to change the encoder to nvenc and cuda respectively.
I know cuda is working since gpuz shows load on the gpu
Both are faster than a CPU only encode.

Can you direct me to these Smokey readings ?

Might shed some light on it for me.....

Also which software was used to produce such results?
(Media espresso was working on a beta I believe, but its not public?)

Cheers,

[deadrats]

http://www.guru3d.com/article/geforce-gtx-680-review/6

the kepler based nvenc is roughly twice as fast as previous cuda based nvenc; i'll look for the other reviews i saw tomorrow, maybe your results are either a result of some idiosyncrasy with media coder or perhaps the target format isn't processor intensive enough to benefit from the new chip, the results i saw were with HD content.

[jagabo]

Has the GPU encoding quality improved at all? Deadrats, this is a serious question, not a dig.

[sharkbite]

Quality is also only a secondary consideration for me - i just want to watch TV shows on the train on the way to work,
that i can't watch the night before due to being totally stuffed....
(but enough about me!)

quality using mediaencoder or freemake is no better or worse for the purpose...

tried media espresso......it sure uses up the system resources!

.....but it takes fer freakin ever to complete...... i am down to around 75 fps using it
(not measured, but it took 30 minutes to complete a 1.5hr show)
and the quality is WOEFUL.

The fastest conversion software on my setup is good ole'' freemake.......
it may only use half the resources, but at least i can effectively double the conversion
speed by doing 2 encodes in parallel....

This means that the GTS250 is going back in the machine, and the GTX650 is going in the kids PC...

(sigh)

[sharkbite]

Update:

Decided to put gts250 in kids machine....
E6750 dual core....

Encoding the same video file in the same way resulted in the same time for output using the same percentage of system resources.

Looks to me like cuda has a speed ceiling no matter what the hardware.....

Encoding the file with CPU only had a vastly different result......
The 1090t is about 3x faster than the e670.......


P.s. also tried xilisoft on the big machine......same output time and quality as free make only its pay ware.....

[jagabo]

That's the general consensus I've seen (though I haven't really kept up with it for the last year): GPU h.264 encoding speed doesn't scale well with the GPU power.

[deadrats]

Has the GPU encoding quality improved at all? Deadrats, this is a serious question, not a dig.

the reviews i've seen don't really delve into IQ, but that's understandable as there are only 2 apps that currently support kepler nvenc, the cyberlink beta and media coder. there have been some passing comments along the lines of "no visual artifacts" when encoding at 720p with a bitrate of 10mb/s, but at 10mb/s + 720p one wouldn't expect to see artifacts with any encoder.

i do know that media coder enables a temporal denoise filter by default if either QS or CUDA or nvenc is chosen and all reports indicate that the output quality is more than acceptable for devices such as ipads and similar such devices.

the sad reality is that consumers really do get bottom of the barrel software, the really good stuff meant for enterprise audiences is just too expensive. there's a guy over at the doom9 forums, one of the mods who is also a video pro and he has stated that the enterprise level software by elemental (the same guys behind badaboom) which includes CUDA powered VC-1, mpeg-2 and h264 encoder (which costs an arm and a leg) and he says that not only is that CUDA powered encoder 2-3 times faster than a xeon based system running x264 but that the quality is also top notch, but those CUDA encoders aren't running on consumer grade gaming cards, they are running on top of the line $4000+ quadro's.

who knows, maybe nvidia is purposely holding back the quality of the encoding consumer level gaming cards are capable of, either with some driver code or something in the micro op code, so as to not cannibalize sales of their high end workstation professional video editing cards.

if that's the case, hardware encoding for the consumer will be relegated to a second class citizen forever.

[sharkbite]

Thanks deadrats...

Just re ran media coder with nvenc.......it seems to not use the gpu at all....

Have you used it, and what settings did you use?


Cheers,

[deadrats]

Thanks deadrats...

Just re ran media coder with nvenc.......it seems to not use the gpu at all....

Have you used it, and what settings did you use?


Cheers,

it's not supposed to use the gpu, using the gpu would imply that the processing is being handled by the CUDA cores, the hardware NVENC is a separate unit, specifically nvidia extended the capabilities of the ASIC decoder chip. download a copy of gpu-z which will allow you to monitor various things about your gpu usage, start an encoding session with NVENC and monitor both windows task manager and gpu-z to see if any significant load is being put either on the cpu or the gpu. if both are near idle and the gpu ram is being loaded up and media coder shows that it's transcoding the file, then you know that NVENC is working properly. hardware NVENC is still new technology and as far as i know there isn't really any type of graphical monitor to show that it's working, the only way you'll know it's doing it's job is when the encoding task is happening but there is no significant load on either the cpu or the gpu.

i think it was in Bruce Almighty that Morgan Freeman (who is playing God) says to Jim Carrey "when you do thinks right, no one will know that you have done anything at all" (i'm paraphrasing here), same thing with NVENC, if it's functioning right you should have any indication of any heavy lifting being done by your pc.

[sharkbite]

Ta matey that's exactly what I have been doing....(buried way up there in post 3)

Using nvenc in media coder has the CPU at 100%.......gpuz shows 0%....
So it looks like its not using the Kepler schmepler at all.....

I would like to believe that there will soon be a "flood" of applications that use
Cuda to its full potential, or nvenc at all.....
But at this stage the gods of software are sorely testing my faith

May I ask - what card do you have, and what s/w do you use?

Cheers,

[jagabo]

I would like to believe that there will soon be a "flood" of applications that use
Cuda to its full potential, or nvenc at all.....

Cuda's been around since 2007. Why would there suddenly be a flood of support?

[sharkbite]

Ah you would have to have seen Bruce almighty to understand the flood reference

[DarrellS]

Has the GPU encoding quality improved at all? Deadrats, this is a serious question, not a dig.

That's what I noticed when I got my NVidia card a couple of years ago. The encodes were a little faster but the quality wasn't as good as using x264.exe at superfast preset which encoded just as fast with at least the same quality. I was hoping that someone would come up with arguments for cudaH264Enc.exe to use with Virtualdub's external encoder feature so I could do comparisons but never found any and gave up.

[deadrats]

May I ask - what card do you have, and what s/w do you use?

i have a gts250, a quadro 400 though lately i've been using the HD4000 built into my 3770k. as soon as i get my tax refund i was going to buy a mid range kepler based gpu and i'm seriously considering putting together a workstation built around said kepler and a 6 module 12 core opteron.

[sharkbite]

Cool.....hope someone comes up with some software that makes it worth the trouble....

All I have now is one that claims to use nvenc but doesn't.......
And the rumour of another which is a beta version that is no longer available.

So....

For all intents and purposes.......at the time of writing this ,nvenc supported video encoding does not exist.....

And cuda is only faster if you have a dodgy CPU to start with



Cheers,

[jagabo]

Has the GPU encoding quality improved at all? Deadrats, this is a serious question, not a dig.

That's what I noticed when I got my NVidia card a couple of years ago. The encodes were a little faster but the quality wasn't as good as using x264.exe at superfast preset which encoded just as fast with at least the same quality.

Yeah, that's why I asked.

[sharkbite]

Further update:

Not expecting a response I emailed nvidia.

At the time of writing this they tell me they are not aware of any applications that use nvenc.


I asked back how did they know it works, and how did they test it?

The response to that was that they don't have an app, and don't know when one will be available.....

Dya reckon we have all been gyped?

[Selur]

Here are some things to remember:

1. there are two things that are called nvenc:
   1. a library that is around for quite some time, which allows to use the encoding chip on NVIDA cards (this one can i.e. be used through the Cuda.exe I wrote quite some time ago; attached it for those you want to use it)
   2. the, with kepler, newly introduced encoder chipset (don't know a freely available command line encoder for this)
2. the normal graphic card has a VPU and a GPU core and for de-/encoding the VPU is used, so measuring the GPU load won't help

(removed attachment, newer 32bit version are attached to a later post,..)

[deadrats]

@Selur

reading through the help file it seems to indicate that this encoder is capable of both gpu accelerated h264 and vc-1, is that correct?

is this something you coded (with the help of that other guy) or is it just a front end for the nvidia reference encoder?

is it open source or is it possible to get a copy of the source?

and lastly, if this is a custom created encoder, are you guys planning on coding a custom h265 encoder as well? it's my understanding that the h265 spec has been designed with massive parallel computing in mind, thus making it very friendly to easy heavy threading and high core counts and as a consequence easy gpu acceleration.

[Selur]

reading through the help file it seems to indicate that this encoder is capable of both gpu accelerated h264 and vc-1, is that correct?

it could until the nvidia folks changed the encoder libraries back in 'October 25, 2010', see: http://docs.nvidia.com/cuda/samples/3_Imaging/cudaEncode/doc/nvcuvenc.pdf

is this something you coded (with the help of that other guy) or is it just a front end for the nvidia reference encoder?

this is a cli to the nvidia reference encoder

is it open source or is it possible to get a copy of the source?

will send you a pm

and lastly, if this is a custom created encoder,...

which is sadly is not,...

Cu Selur

[DarrellS]

Selur, are you familiar with Virtualdub's external encoder? Is there a command argument to make the Cuda.exe work with it?

I tried before using the cudaH264Enc.exe CLI with an argument that I found in a thread about it but Virtualdub couldn't use it. Here was the argument that I tried to use...

" -i \"$(SourceFile)\" -o \"$(tempvideofile)\" -profile 0 -preset -1 -idrp 15 -qp 25 -qpp 28 -qpb 31 -deblock -pinterval 3 -rc 1 -abit $(VideoBitrate)",

I also tried Cuda.exe with....

"--input-res %(width)x%(height) --fps %(fpsnum)/%(fpsden) --profile main -o \"%(tempvideofile)\" -",

Both encoders seem to need the input file but the file is already decompressed in Virtualdub. It just needs a command to demux the raw video and encode it to .264.

x264 command is...

"--crf 21 --preset medium --tune film --demuxer raw --input-csp i420 --input-res %(width)x%(height) --fps %(fpsnum)/%(fpsden) -o \"%(tempvideofile)\" -",

[Selur]

I know of Virtual Dubs external encoder support, but I never tried it.
After a quick look at http://forums.virtualdub.org/index.php?act=ST&f=3&t=18840 here are command arguments that my Cuda.exe should work:

--resolution %(width)x%(height) --input - --sar 1x1 --format IYUV --control_mode cq --fps %(fps) --profile high --level auto --offload partial --measure FPS --showFrameStats 100 --deinterlace true --frame_typ frame --pframe_dist 1 --gop_max 250 --dynamicGOP true --pquant_min 20 --bquant_min 24 --iquant_min 20 --deblock true --cavlc false --nal_typ auto --sps_pps false --slices auto --output "%(tempvideofile)"

there is no equivalent to x264s call since the nvuenc library doesn't have something like: crf, preset medium or tune film.

----
No clue where you got the cudaH264Enc.exe from or what it's parameters are,...

Cu Selur

[DarrellS]

No. It didn't work. I'll try tweaking it later and see if I can get it going or post it on the Virtualdub site and see if the guys over there can get it going.

I must've gotten the cudaH264Enc.exe from Mediacoder since I believe they were the only one with a Cuda encoder at the time.

I'd love to get one of these Cuda encoders working with Virtualdub since it's the only editor that I use. It's come a long way since the first version that I used back in 2000.