gtx 1050 vs 1060 vs 1070 in encoding?

[dwaynek]

do these 3 cards perform the same when encoding video?

[raffriff42]

Yes I think so. It looks like there's no NVENC performance difference within a given GPU generation.

1. https://en.wikipedia.org/wiki/Nvidia_NVENC
2. https://developer.nvidia.com/nvidia-video-codec-sdk#NVENCPerf
3. https://www.wowza.com/forums/showthread.php?41278-NVIDIA-card-NVENC-performance-spec&p...677#post150677

[pandy]

There are some differences between chips used on cards - newer chip offer improvements and bug fixes. so 1050 may have better NVEnc than older cards... and this doesn't mean that raffriff42 is wrong - simply some of those things are not entirely documented by NVidia... (some pieces of information like new bitrate allocation modes or support for higher bitdepths):
https://www.tapatalk.com/topic/17046-videohelp-com/381493-nvidia-pascal-hevc
Silicone is the same as software - it evolve every iteration (every new silicone cut).

[raffriff42]

Good points, pandy.

Also, from your link:

But I also read that in encoding, a lot of VRAM is used and more memory bandwidth is better?

This is something I wondered about as well. It sounds logical, right? But until somebody publishes benchmarks, we won't know. I've been looking for NVENC benchmarks listed by GPU, but haven't found any, so far.

[pandy]

Good points, pandy.

Also, from your link:

But I also read that in encoding, a lot of VRAM is used and more memory bandwidth is better?

This is something I wondered about as well. It sounds logical, right? But until somebody publishes benchmarks, we won't know. I've been looking for NVENC benchmarks listed by GPU, but haven't found any, so far.

NVidia release only limited set of data - NVenc is black box - there is no clear dependency between memory subsystem clock and overall GPU (3D) clock - perhaps NVenc is clocked differently... It is not even clear how to test NVenc capabilities as NVenc API expose only limited aspect for this block.

[sophisticles]

I'm currently using nvenc via ffmpeg on Ubuntu 16.04 LTS, with a GTX 960 and 367.57 drivers, encoding the y4m version of Tears of Steel 4k DCI source to 4k DCI hevc in a mp4 container at 18mb/s with the following command line:

time ffmpeg -y -i tearsofsteel-4k.y4m -vcodec hevc_nvenc -preset slow -rc-lookahead 10 -b:v 18M ToS.mp4

time ffmpeg -y -i tearsofsteel-4k.y4m -vcodec h264_nvenc -preset slow -rc-lookahead 10 -nonref_p 1 -no-scenecut 0 -b_adapt 1 -spatial-aq 1 -aq-strength 15 -temporal-aq 1 -b:v 18M ToS_h264.mp4

It's encoding in 1.2x real time which works out to about 29 fps for hevc, 1.44x (35fps) for h264; using the Nvidia X Server control panel to monitor what the card was doing, here's the essentials:

During HEVC Encoding
Total Memory 2048 MB
Total Dedicated Memory 1988 MB
Total Used Dedicated Memory 650 MB (33%)
PCIe Bandwidth about 44%

During H264 Encoding
Total Memory 2048 MB
Total Dedicated Memory 1988 MB
Total Used Dedicated Memory 732 MB (37%)
PCIe Bandwidth about 44%

GPU Utilization 66%
Video Engine Utilization 60%

I'm thinking about buying a Pascal based card, like a 1050, though it's technically a slower card from a gaming perspective everything I have read says that it's much better from an encoding perspective.

If I enable hardware decode with a compressed source the GPU Utilization and Video Engine Utilization approach 100%.

[pandy]

1050Ti should be comparable to 960 http://gpuboss.com/gpus/GeForce-GTX-960-vs-GeForce-GTX-1050-Ti