New AMD Ryzen Threadripper Editing Machine

[nctnico]

When using a GPU for computational tasks you may run into a different problem: insufficient cooling. During my cryptocoin mining experiment it turned out the cooling solution on the GPU I was using was only enough to get roughly 1/3 of the performance from the GPU.

[hans]

Similarly; new Intel CPU's are getting less specific in their "turbo" all-core clocks (compared to base), as it heavily depends on the workload (as we all know I think). This is also yet another reason to pay the few 10$ premium for a K part, as they have higher spec'ed turbo and base clocks, which helps with minimum guaranteed performance..

In particular, "AVX workloads" (advanved vector instructions) are used a lot in CPU video encoding, and are particularly high power. Especially Intel are quite strong with their quite recent AVX2 instruction sets, which 1st gen AMD Threadripper and Ryzen are lacking. If you look up H265 encoding benchmarks you'll see Intel pull ahead quite well, and I think there is still some deficiency in H264 as well; https://uk.hardware.info/reviews/7585/7/intel-core-i9-7980xe--7960x-review-intel-back-in-the-lead-with-18-cores-benchmarks-video--and-audio-encoding-x264-x265-and-flac

Of course bang/buck ratio is a different story.

[EEVblog]

FWIW some benchmarks for Vegas Pro 15 on the 2 minute project, using MAGIX AVC/AAC MP4 codec Internet HD 1080p 50 fps:

o 1'08" Ryzen 1800x oc 4GHz (8c/16t) NVENC GTX 1080
o 2'02" Dual Xeon E5-2696v2 (24c/48t) NVENC GTX 1070
o 2'18" i7-8700k (6c/12t) NVENC GTX 1050 ti
o 2'39" i7-8700k (6c/12t) Intel QSV (iGPU acceleration)
o 2'52" Dual Xeon E5-2696v2 (24c/48t) CPU only
o 3'14" Ryzen 1800x oc 4GHz (8c/16t) CPU only
o 4'07" i7-8700k (6c/12t) CPU only

These render times seem pretty slow to me for h.264 1080p 50fps, but I am not a regular Vegas user: doing the same two minute AVC render on PowerDirector 16 with HVENC on the GTX 1070 took only 0'23", and it was a higher bitrate (27Mbps). Even upscaling it to 2160p rendered in 1'24".

Seems all over the shop. My dual Xeon with a GTX 1050 has done both 1:10 and 1:35 for the same thing. Crazy.

[nctnico]

FWIW some benchmarks for Vegas Pro 15 on the 2 minute project, using MAGIX AVC/AAC MP4 codec Internet HD 1080p 50 fps:

o 1'08" Ryzen 1800x oc 4GHz (8c/16t) NVENC GTX 1080
o 2'02" Dual Xeon E5-2696v2 (24c/48t) NVENC GTX 1070
o 2'18" i7-8700k (6c/12t) NVENC GTX 1050 ti
o 2'39" i7-8700k (6c/12t) Intel QSV (iGPU acceleration)
o 2'52" Dual Xeon E5-2696v2 (24c/48t) CPU only
o 3'14" Ryzen 1800x oc 4GHz (8c/16t) CPU only
o 4'07" i7-8700k (6c/12t) CPU only

These render times seem pretty slow to me for h.264 1080p 50fps, but I am not a regular Vegas user: doing the same two minute AVC render on PowerDirector 16 with HVENC on the GTX 1070 took only 0'23", and it was a higher bitrate (27Mbps). Even upscaling it to 2160p rendered in 1'24".

Seems all over the shop. My dual Xeon with a GTX 1050 has done both 1:10 and 1:35 for the same thing. Crazy.

As I wrote above: the cooling on a graphics card may not be sufficient to get maximum performance. Once I rigged my cryptomining setup with a blower on the GPU and airflow ducts so cold air got sucked in and hot air got blown out it could deliver twice the performance so the difference is not just a few percent.

[Monkeh]

The GTX 1050 is quite a low power card, and the hardware encoding doesn't even use the main GPU hardware. Thermals are really not a problem.

[amspire]

The GTX 1050 and 1060 families have one NVENC engine. The GTX 1070 and 1080 families have two NVENC engines. The GTX1030 uses a GP108 chip and I am not sure if it has any NVENC engine at all.

As I have mentioned before, these engines are capable of about 350 frames per second each of high quality 1080P H.264 and so the speed is dependant on the speed of the encoding streams feeding the engines. For 1080P high quality, it seems the optimal number of streams need is about 9 for two engines and 4 or 5 for one engine. That is based on the nVidia data on the Quadro cards that use the same chips as the GTX family and that are allowed to run unlimited numbers of encoding streams.

Licensing issues means that the non-Quadro cards are limited to two streams so that is the bottleneck. Potentially, you could get the similar NVENC speeds from a 1060 or a 1080. When you allow Vegas to use CUDA rendering, it uses it for the special effects in the editing, so it really does help if you have added a lot of effects. If not, CUDA rendering doesn't speed things up at all.

The GTX 1050 chip seems to only be capable of two HQ encoding streams but if the streams are efficient enough, it still could possibly perform almost as well as a GTX1080. It all comes down to the software feeding the cards.

So if you have a GTX1070/1080, you have a card with an encoding engine capable of 700 1080P High Quality H.264 frames a second,  you are doing very well if you can feed it quick enough to get 200 frames per second.

If you are the proud owner of a Titan, then you have 3 NVENC engines but you can still only run 2 streams. I guess it is always handy to have a spare.

I don't have a fast modern CPU right now but I did find MediaCoder seems to be a fairly efficient transcoder that includes NVENC. You are meant to have to pay for GPU encoding, but NVENC worked fine on my free version. On my old system, it did seem to use the CPU processors more efficiently then Vegas. It is possibly a good way to get the real NVENC of a CPU/NVENC speed. It used NVENC 7.0 and the latest SDK is NVENC 8.0. You should have nVidia drivers 378.66 or newer (Windows) or 378.13 or newer (Linux).

http://www.mediacoderhq.com/download.htm

[Marco]

I still fail to see the logic in using a hardware encoder. A couple minutes of background activity for better quality and/or less bandwidth for thousands of viewers.

[Mr. Scram]

I still fail to see the logic in using a hardware encoder. A couple minutes of background activity for better quality and/or less bandwidth for thousands of viewers.

It's ending up on Youtube, so any arguments about quality are pretty much moot.

[amspire]

I still fail to see the logic in using a hardware encoder. A couple minutes of background activity for better quality and/or less bandwidth for thousands of viewers.

The problem here is that Dave's dual Xeon is strangely slow with both software and hardware rendering and the rendering times also seem erratic. I think he would be happy to solve the slowdown problem with either CPU or GPU rendering.

[kaz911]

I just did a test with DaVinci Resolve compared to Vegas Pro 15 - and Resolve renders about 30-40% faster on the same PC (NVidia M4000 / Xeon Mobile 1505v3) - and more or less same settings. (1 main video - 2nd picture in picture video 1080p)

Resolve pushes CPU to 100% where Vegas only pushes to about 60-70% - GPU load not very high on either(15-20%) -  Vegas pushes on the Encode/Decode part of the GPU where Resolve seem to use the 3D part according to Task Manager Win10.

[Electro Detective]

Which version/Build numbers of both programs, and what hardware and operating sytem? 

[kaz911]

Which version/Build numbers of both programs, and what hardware and operating sytem? 

Win 10 latest crappy update
ThinkPad P70 - External USB 3 SSD (about 500/500 read write speed)

Resolve : latest v14 downloaded 2 days ago
Vegas Pro v15 - build 261 (latest)

Output 1080p / 25fps - mp4 - 20mbps max on both. Vegas using NVENC - Resolve using whatever it wants (it makes its own decisions)

Only "effect" - Picture in picture - rotate 180 deg  - 720p stream on top of 1080p with 720p scaled to about 1/6 lower left corner. In Vegas it is an built in Effect - in Resolve it is a "Inspector/Transform"

Resolve - Lots less crashes

Original media prepared through PluralEyes for Sync (you load a bunch of video files and it auto sync's them based on sound) - in Vegas you run it FROM Vegas. With Resolve you do it manually and export as Premier XML then import into Resolve. PluralEyes did about 9 hours of recording X 2 cameras - and synced them up in about 6 minutes. And the 720p camera is a terrible camera with really bad sound.