NVENC Streams Transcode

[pir8radio 1294]

I want to offload some video transcoding to a GPU, my xeon CPU doesn't support quicksync so I must go this route...  

 

Looking at this list https://developer.nvidia.com/video-encode-decode-gpu-support-matrix   it looks like the P2000 might be a good test GPU, it says "unlimited" streams.   Does anyone have any NVENC experience with any of these processors, I don't want to hard limit my server to 2 or 4 streams if possible.   I don't want to spend money without getting some input from you guys first.    I have a chance to pick up a Tesla P4 for a good price, but I would rather test on the P2000 first.    I recall some earlier nvidia cards only supporting 2 streams, but this list says differently..   Any input would be greatly appreciated. 

 

@@Waldonnis might have some input? 

Edited August 6, 2017 by pir8radio

[pir8radio 1294]

So to be clear, in the above post, I'm just asking if Unlimited streams = no hard limit in emby, as to how many NVENC concurrent transcode sessions. 

 

I understand that there is a limit to what the card can handle. 

[Luke 37367]

There is no hard limit in Emby.

[RanmaCanada 362]

Only the consumer grade cards are limited to 2 streams.  All Quadro based cards are "unlimited", well as long as they are above the 2000 line, but in reality are limited based on the speed of the processor that is on the chip.  As it's kepler, it's going to be very limited in the amount of streams it can actually handle, specially at 1080p.  You might be lucky if it can handle 3.

[pir8radio 1294]

Only the consumer grade cards are limited to 2 streams. All Quadro based cards are "unlimited", well as long as they are above the 2000 line, but in reality are limited based on the speed of the processor that is on the chip. As it's kepler, it's going to be very limited in the amount of streams it can actually handle, specially at 1080p. You might be lucky if it can handle 3.

This is the info I was looking for. This is a 'P'2000 which is a pascal processor. How do you think that card will perform? Otherwise I may go straight for the p4 or if I can find one cheap ha, a p40.

 

http://www.pny.com/nvidia-quadro-p2000

[RanmaCanada 362]

I had the information posted in the beta thread I believe.  It used to be that all quadro cards were unlimited but nvidia changed it to anything 2000 or higher.

My bad I thought it was a kepler you were getting, not a Pascal.  

 

A general rule of thumb is that all Quadro's below X2000 (X = K, M, P) and all GeForce cards have limited encode sessions.

https://devtalk.nvid...-session-limit/

 

Which didn't used to be the case.  https://devtalk.nvid...venc-sessions-/  Nvidia changed things and screwed everyone over.

 

If it's a P2000 that is basically a 1060.  Sorry for the confusion, I'm a moron at times

 

card                             NVDEC NVENC H264 NVENC H265 CUDA DEINTERLACE

QUADROM4000:        1250           2300*     1200*       4000 

GTX 960:                    1800        1800          900         3000

GTX 1060:                  2600         2600        1800          4000

GTX 1070:                  2600        2600        1800            5000

GTX 1080:                  2600        5200*        2600*        10000

Encoding and decoding is normalized to 720x576 resolution and units are FPS!

If you want to know speed for:
HD (1280x720) - divide all by 2
FHD (1920x1080) - divide all by 4

For example encoding on GTX 1070 in FHD quality to H264 will run 650 FPS

* those cards have 2 NVENC engines, so speed for only one thread will be half

https://devtalk.nvidia.com/default/topic/987460/nvdec-cuda-nvenc-speed-comparison/

 

So you're looking at a total of 650 fps (2600/4) split between all your encodes.  So you could theoretically do 21 transcodes in real time at 30 fps, providing you have the ram for it and that the card could keep up (ram in card etc) and your hard drives, etc etc.  

Edited August 7, 2017 by RanmaCanada

[pir8radio 1294]

I had the information posted in the beta thread I believe.  It used to be that all quadro cards were unlimited but nvidia changed it to anything 2000 or higher.

My bad I thought it was a kepler you were getting, not a Pascal.  

 

A general rule of thumb is that all Quadro's below X2000 (X = K, M, P) and all GeForce cards have limited encode sessions.

https://devtalk.nvid...-session-limit/

 

Which didn't used to be the case.  https://devtalk.nvid...venc-sessions-/  Nvidia changed things and screwed everyone over.

 

If it's a P2000 that is basically a 1060.  Sorry for the confusion, I'm a moron at times

 

card                             NVDEC NVENC H264 NVENC H265 CUDA DEINTERLACE

QUADROM4000:        1250           2300*     1200*       4000 

GTX 960:                    1800        1800          900         3000

GTX 1060:                  2600         2600        1800          4000

GTX 1070:                  2600        2600        1800            5000

GTX 1080:                  2600        5200*        2600*        10000

Encoding and decoding is normalized to 720x576 resolution and units are FPS!

 

If you want to know speed for:

HD (1280x720) - divide all by 2

FHD (1920x1080) - divide all by 4

 

For example encoding on GTX 1070 in FHD quality to H264 will run 650 FPS

 

* those cards have 2 NVENC engines, so speed for only one thread will be half

https://devtalk.nvidia.com/default/topic/987460/nvdec-cuda-nvenc-speed-comparison/

 

So you're looking at a total of 650 fps (2600/4) split between all your encodes.  So you could theoretically do 21 transcodes in real time at 30 fps, providing you have the ram for it and that the card could keep up (ram in card etc) and your hard drives, etc etc.  

 

Very good info, Thanks.   Hopefully what you say is true!   I guess Ill have to spend the cash and get a P2000, put it to the test.        My xeon does well, but it can easily peg what cores I allow it to eat, and all I could do is add a second xeon but for the price of a second CPU I could get a few P2000's and get more bang for my buck.....  At least that is my hope.  My drive array will be the bottleneck.  But I'm fine with that...      Ill have to purchase and report back...   Thanks all for the info.    I'm still open to additional info if anyone wants to post. 

[RanmaCanada 362]

You might be better off to switch everything over to quicksync.  Apparently the E3 series with iGPU support it.  E3-1275 v3 scores almost 10 000 passmark and looks to be had for pretty cheap.  It's socket 1150 so it would mean new hardware sadly, but it would be an upgrade!

[Balthazar2k4 0]

Very good info, Thanks.   Hopefully what you say is true!   I guess Ill have to spend the cash and get a P2000, put it to the test.        My xeon does well, but it can easily peg what cores I allow it to eat, and all I could do is add a second xeon but for the price of a second CPU I could get a few P2000's and get more bang for my buck.....  At least that is my hope.  My drive array will be the bottleneck.  But I'm fine with that...      Ill have to purchase and report back...   Thanks all for the info.    I'm still open to additional info if anyone wants to post. 

 

I am looking at the P2000 as well should my VAAPI on the G4600 fall through.  Did you get one and what has your experience been?

[Luke 37367]

The next release of Emby Server for Linux x64 will include NVENC support out of the box so all you need to do is enable it. 

 

Nvidia Cuda and OpenCL drivers will need to be installed first. These cannot be freely distributed by us so you will need to refer to your respective distro to learn how to install them.

 

Thanks !