Movie Stud15 - NVEC encoding awful encoding performance, not using GPU

whatdoido wrote on 4/3/2018, 2:00 PM

One of the reasons I upgrade from MSP13 to 15 was because of claims of hardware accelerated NVENC and the positive links like this:

https://www.vegascreativesoftware.info/us/forum/any-improvement-in-rendering-speed-with-vegas-pro-15--108314/#ca667331

However from my current experience, version 15 barely improves on GPU encoding. So I want to know if we've been sold snake oil or it's user or error "work in progress"!!!

My system is old, an 8x core i7 870 12GB RAM Win7x64 machine, but the GPU is a GTX1060 3GB card.

For a simple comparison to demonstrate a point, I took a 6min 23sec GoPro file that is 2.7k @ 24fps and I want to downsize to 1080, no edits added to timeline. I run this through MSP15 MagixAVC/AAC MP4 Internet HD1080 NVEC (CBR and high performance) and MSP15 took 6 MINUTES. The CPU is at about 70-90% but the GPU (through GPUz) is only 13-16% utilised

As a programmer in my former life it is very clear to me the GPU is not being used correctly.

To prove that my system is not the bottleneck, I dual boot to linux and run a NVEC enabled ffmpeg. I downsize the same GoPro file from the same harddisk and write the output to the same output disk as used by MSP.

.... and ffmpeg it took 31 SECONDS !!! Yes MSP15 vs ffmpeg/linux is not direct comparison and the encoding options are not exactly the same but, COME ON, the new NVEC Movie Studio is clearly not working as expected.

ffmpeg was using full h/w transcode, using the GPU memory and not system memory. When I force ffmpeg to h/w decode/encode but with system memory, it takes a little over 2mins 30seconds (still better than MSP) and CPU only takes closer to 10mins.

MAGIX - Comments/thoughts please.

--- ffmpeg command for reference

# full h/w decode/encode using gfx card memory - remove "-hwaccel cuvid" to disable

ffmpeg -y -hwaccel cuvid -c:v h264_cuvid -i gopro.mp4 -vf scale_npp=-1:1080 -c:a copy -c:v h264_nvenc /export/public/foo.mp4

Back to post

Comments

ggrussell wrote on 4/6/2018, 2:34 PM

Downloaded your file. First I wouldn't consider this subject complex for rendering. Lots of same colors although there was plenty of motion. Using the same settings posted by JN_, my results:
Mainconcept
CPU Core- 1556.5MHz
Memory - 961MB
GPU Load - 6-8%
Video engine - 0%
2min 33 sec

NVENC
CPU Core - 1949MHz (max for my GTX1060)
Memory - 1212MB
GPU Load - 15-26%
Video Engine - mostly 0% but there were several (maybe 20 guessing) frames it jumped to 70%
rendering time 32 sec

Just for comparison, I used my other NLE. Tried to get the output profile as close to those posted here. File rendered in less than 6 seconds! The other NLE does use the GTX1060 video engine.

whatdoido wrote on 4/8/2018, 6:53 AM

Thanks - yes, the test video was deliberately 'simple'.

Now you 32 seconds with NVENC is still better than my > 1m encoding with MSP so without MAGIX responding (I've raised a support ticket but not holding my breathe) we don't know why this software can't do what it promised.

And as I've mentioned earlier in this thread, I know the same physical system can process the same file (using NVENC enabled ffmpeg) in about 7 seconds - this matches your experience with the alternative NLE (do you mind sharing the name?) - so we know the GTX1060 can do this ..

We can therefore, based on experiences here, conclude MAGIX MSP/VP does NOT fully support/utilise NVENC/Nvidia GPU accelerated encoding even though it claims this in their product pages and this should be made public and buyer-beware!

Former user wrote on 4/8/2018, 9:39 AM

@whatdoido

Just re-reading this thread and trying to get a conclusion...

“TLDR; no surprises

ffmpeg with NVENC full h/w dec/enc, ~7 (SEVEN) seconds (best case)

ffmpeg with NVENC h/w enc using system ram, ~32 seconds

MSP15 NVENC, 59 (FIFTY NINE) seconds

The NVENC options are matched as best I understand them between the presets available with ffmpeg/nvenc.”

When you do this what are the file sizes and data rates for the ffmpeg and Magix output files? Are they similar? re: the 7 second file.

As I previously mentioned I get error messages running the static ffmpeg in Windows. I see now that you dual booted to Linux, I don’t want that hassle, but if ffmpeg can be run in Windows with hardware acceleration from a static build i’ll give it a go. I got this error message .. no such filter, “scale_npp”.

Former user wrote on 4/8/2018, 1:45 PM

Ok I couldn’t get your gopro file sample to scale to fhd in my static ffmpeg so I used a new 22 second long 4K 25 fps file sample and rendered it out to 1920x1080 for tests.

I tested on my laptop. I ran ffmpeg with and without hardware acceleration.

I ran VP15 with and without Nvenc (forgot to use MSP15, but we’ve already proved their encoding is the same).

No matter what settings I put into the ffmpeg command line for HW acceleration it always produced a max of 2,212 kb/s, 6.15 mb size. The original 22s file is 255 mb and has a data rate of 95,586 kbps.

So I changed Magix and ffmpeg (ffmpeg non HW accelerated) settings to roughly give output that matched this 2.2 mb/s data rate and 6.15 mb size on disc.

ffmpeg with HW acceleration... 0:21s

ffmpeg without HW acceleration... 0:32s

Magix with Nvenc ... 23s

Magix without Nvenc ... 32s

Former user wrote on 4/8/2018, 3:50 PM

But! This was using only -cuvid, no -hwaccel prefixed (because I couldn’t get the scaling to work when I used -hwaccel).

So I used -hwaccel and just converted your gopro clip to the same output resolution, no scaling.

ffmpeg with HW acceleration... 0:6s

ffmpeg without HW acceleration... 1:37s

Magix with Nvenc ... 0:43s

Magix without Nvenc ... 2:23

All matched to similar size on disk, about 16.2 mb and similar data rate, about 2130 kbps.

Former user wrote on 4/9/2018, 8:14 AM

Update: I found a ffmpeg setting to modify data rate .. -rc constqp -qp 21

The value 21 is roughly what gets me around 20mb/s for your test file, range is I think 1 to 51.

So I ran ffmpeg and Magix again to both give approximately 20 mb/s data rate and a size on disc range of 147 to 159 mb. No scaling applied.

ffmpeg with HW acceleration... 0:6s-0:7s

ffmpeg without HW acceleration... 2:58s

Magix with Nvenc ... 0:43s

Magix without Nvenc ... 4:53s

Note: my previous .. ffmpeg without HW acceleration... 1:37s. This might be HW acc. using system memory.

Zingfharn wrote on 6/10/2018, 7:17 PM

Was this ever resolved? I just upgraded based on the GPU acceleration support, but it seems to be doing nothing.

Looking at GPU-Z while encoding, usage on a GTX1080 is ~20% while CPU is 100% and render times don't seem terribly improved. I'm using the MAGIX codec, VBR 25 av 28 max, VBR and default, NV encoding.

Am I missing something obvious? GPU processing shows the correct processor.

mmcswnavy24 wrote on 6/10/2018, 9:33 PM

Okay, decided to try this out on my machine.

PC Specs: i7-5820K (Haswell-E) stock 3.3 GHz, 32GB Corsair LPX @ 2667MTs, Gigabyte WindForce GTX 1080 (stock) with driver 398.11, OS on Intel 600p NVME and Magix MSP15 Build 116, Win 10 64-Bit Build 17134.48, Render drive Crucial M200 512GB SSD, Source drive WD Black 4TB SATA III.

So, no changes to file from download, using the settings for render as Magix AVC Internet HD 1080 (NVENC), with setting at 23.976 PAR at 1.000 Encode NV Encoder Preset Default RC Mode VBR Bit Rates 24/20 (Just like what was posted - so had to change original Render setting of 29.97 to 23.976 and save).

Wow, amazing...23 second render time. GPU-Z showed 24% usage, while Hardware Monitor showed it at 46%.

Not sure which GoPro camera this came from, but might re-try all of my 4k footage from my Hero 5 Black and test. I know from using the footage before, from advice on Derek Moran's "Movie Studio Zen" YouTube channel, I would convert the footage to AVI via the GoPro Studio software. Maybe I will buy the Hero 6 Black and try out that newer HEVC 4K footage!

Oh, and the i7-870 "Lynnfield" processor is 4 Core/8 Threads. Yeah, was a good processor in its day. Glad I upgraded to the X58 platform, to X79, and now on X99 (two systems). Now, to downgrade to only one system but still trying to decide on either ThreadRipper X399 with 1950X or X299 with either i9-7900 or i9-7920. Tough choice.