4K Benchmarking Roundup

Howard-Vigorita wrote on 8/3/2019, 9:14 PM

Here are some comparative benchmarks that include Vegas Pro v16.424 and Vegas Pro 17.284. 

http://rtpress.com/roundup.htm

(UPDATE: also see HEVC 2020 thread)

Started working on this before last Christmas when I was considering the purchase of a 4K camera.  In the collaborative I often work with, a couple members have been showing up for shoots with xc10 and bm 4k production cameras and dumping their footage on me.  So I’ve been accumulating a number of mixed-resolution projects under my belt; with my own xf305 and jvc hd’s on other tracks.  My benchmarking objective was to assemble a Vegas system capable of handling projects with more if not all 4K multi-cam tracks.  So my 1st step was to put together an i9-9900k system this past Spring with Radeon VII video.

My Christmas 2018 effort included transcoding the main Sony Red Car footage (clip#3) from its original 1920x1080 50 mpbs mxf format to suitably upscaled 4K.  Studying online sample clips and resolving not to get a camera with a lesser mbps output than my Canon, I transcoded the original 4:30 min mxf to 100 mbps Prores using ffmpeg, matching file size and other parameters of the samples I found, as best I could.  Clips #1 and #2 of the original Red Car are actually just compressed versions of clip#3 squeezed down to around 24 mbps and 13 mbps.  My cohorts seem to think a lowest useable 4K compressed limit might be 20 mbps so we went with that using Mainconcept and AMD VCE encoders on the original mxf. That coincidentally seems to match the bit rate that Magix uses for proxy files, though we didn’t think of looking at that at the time.

I’ve learned allot over the years about how to use Vegas by studying the original Red Car project.  Like how to put a proxy or different camera angle in a box or overlay a track with it faded.  I use that to good effect in concert and theatrical musical videos with the players hands or faces.  And observed that if used in a window or faded like a ghost, a compressed version of the clip gives great timeline performance… and might even be left in the final render without any visible quality loss.  The 4K Red Car tries to follow that learning.  I’ve also included  a 4K Red Car Torture version…  the only difference being that in the Torture project, clip#3 is mapped to both clips#1 and #2.  This is meant to measure what happens when 100 mbps 4K clips are used on all the tracks… keeping in mind that many video projects might not actually play two 4K clips on top of one another for long stretches. But crossfades between cameras do add up.

Mine is not the 1st 4K Red Car effort here.  At least 2 others lead the way with earlier versions of Vegas:

https://www.vegascreativesoftware.info/us/forum/4k-scs-benchmark--99009/

https://www.vegascreativesoftware.info/us/forum/v15-4k-render-benchmarks--108438/

You might have seen my mention of this effort in a couple of other threads this past Spring. It seems to have inspired a new community effort called the Sample Project which I have followed with interest.  You might notice that I admired it so much, I shamelessly ripped off their spread sheet presentation, adjusting column headers a little for clarity, and being a little more fastidious in listing my encode mode.  I encourage everyone to check out the Benchmarking thread, and if you run your own timings for it, be sure to report them there:

https://www.vegascreativesoftware.info/us/forum/benchmarking--116285/

I’ve studied the Sample Project and it’s quite different in approach from the Red Car.  For one, it makes no effort to resemble a real world video project shot with a 4K camera. It uses a single mp4 video track, but at around 10 mbps, that’s kind of window dressing.  The real action is in 4 tracks that play concurrently with the mp4 running a media generator called the Vegas Noise Texture. It basically uses a noise generator to create a simulated video load on the fly.  So the benchmark actually measures the efficiency of Vegas both creating the load and processing it.  The 2 loads could be separated by muting all the other clips, rendering each track having noise generator clips, then substituting rendered output back into the project and alternately muting original and substitute clips.  I leave that as an exercise for others.  To the extent that a media generator resembles a 4K camera clip, it would model such clips being played concurrently on 4 tracks.  Although that might not be common for real world 4K camera projects, the big advantage of the Sample Project benchmark is it’s relatively tiny to download and is a good yardstick of relative cpu and gpu performance under stress. So long as measurements are taken and presented in a way that does not comingle both cpu and gpu in all of its measurements. Assuming you want to use this data to make informed decisions on CPU’s independent of gpu purchases.

In all of my own timings I have taken care not to conflate combined CPU and GPU performance as reported cpu-only timings.  Generally speaking, any gpu performance measurement will use a mix of cpu and gpu… gpu-enabled encoders, however, vary in how and when they choose one over the other.  Which is clearly reflected in observed timing.  But a true CPU-only measurement is required as a yardstick if you want to estimate the GPU contribution of a particular encoder.  I was able to do that for both Mainconcept and AMD VCE encoders because they do not disappear when I disable GPU in Vegas video settings.  When I tested physically removing the GPU, AMD VCE disappeared as a render option but I confirmed that timings for Mainconcept were identical to disabling gpu in Vegas. So I feel reasonably confident that when I only disable GPU in Vegas, it’s not being used when I run AMD VCE which yields much lengthier timings.  QSV is different, however.  I can still run the QSV encoder even though I might not select Intel as the Vegas GPU… and it runs even faster.  Being that it’s an internl GPU that is tightly bound to my Intel CPU and motherboard chipset, I suspect the CPU might be handing off operations to the iGPU without regard to what Vegas asks. So all my QSV timings are with iGPU enabled in bios and all my other timings are with iGPU disabled in bios… in which case QSV disappears as a Vegas render option and I cannot get a CPU-only render with that encoder and view it as a GPU/CPU-only encoder.

Note that I’ve made no effort to assess relative render quality.  Without honest to goodness high quality 4K camera footage, I don’t think that’s possible.

If you want to run your own timings yourself, get the project here:

https://drive.google.com/open?id=1TK2lyFZSlzU9JQ8xO1mvoj-zl9tyj2lc

I’ll try and leave it there for Googles of years, or their lifetime, whichever is shorter . 

Feel free to post your own 4K Red Car timings here. But I’m not inclined to add any but my own to my spreadsheet.  I do have at least 2 more systems of my own that I hope to add… a Microsoft Surface Book and an Intel NUC.  I do, have 4 Vegas licenses to go around, now that I jumped on the 16+17 deal!  But it’ll take a little juggling.  Speaking of 17… expect I’ll be doing it again sometime soon.

Just remember that benchmarking is mostly a waiting game.  During which time reading the refraction patterns in a glass of fine whiskey on the rocks helps pass the time.  Then when all is said and done, you can proudly proclaim, “I Benchmark, and I know things.”

Comments

Howard-Vigorita wrote on 8/3/2019, 10:39 PM

Vegas 16 build 424... Here's a closeup of the i9-9900k timings:

Benchmark timings for the e5-1650v4 Xeon:

Benchmark timings for the i7-980X:

One curious trend I noticed is how high the gpu performance is in all the Sample Project UHD renders compared to the Red Car Torture test. But how the tables turn without the gpu. Seems like maybe the gpu processes noise texture content differently from camera-shot video content.

Howard-Vigorita wrote on 8/6/2019, 3:17 AM

Just added timings for the Intel Hades Canyon NUC:

I've also added a column for a 4K proxy play-ability rating. I arrived at the number by replacing the torture test clips with proxies which Vegas creates when the feature is turned on in Video settings, then tried to play it through with various settings from Best-Full on down. If I can play at 29.97 fps at Best-Full, it gets a 16, Best-Half=15, and so on down to 00 if the frame rate could not be maintained at any setting.

To see how previous machines I posted did, check out the new link:
http://www.rtpress.com/roundup.htm

Howard-Vigorita wrote on 8/7/2019, 5:00 PM

New timings for the 9900k cpu with a Radeon VII for both Vegas 16 Build 424 an Vegas 17 (trial) Build 284:

Hard to distinguish between VP17 and VP16 performance on this system because it's blazing fast no matter what its running.

Web link also updated: http://www.rtpress.com/roundup.htm

fr0sty wrote on 8/7/2019, 5:18 PM

That's a very significant jump from version to version.

Systems:

Desktop

AMD Ryzen 7 1800x 8 core 16 thread at stock speed

64GB 3000mhz DDR4

Geforce RTX 3090

Windows 10

Laptop:

ASUS Zenbook Pro Duo 32GB (9980HK CPU, RTX 2060 GPU, dual 4K touch screens, main one OLED HDR)

Howard-Vigorita wrote on 8/9/2019, 11:50 PM

Vegas 17 vrs 16 benchmark timings for the i7-980X / rx480.

One odd thing I noticed doing the Playback test with Red Car 4K proxies is that both VP17 and VP16 topped out with a setting of Preview-Full on the preview screen. But VP17 seemed smoother. Guess the numbers don't catch all the nuances.

fr0sty wrote on 8/10/2019, 12:00 AM

I see, I was reading the chart wrong, I assumed the ones in white were 16.

There's one thing you need to specify, though... which decoder are you using, or are you using one? Intel chips, some at least, support quicksync, so go into your file i/o tab and see what it is using there. That can make a big difference in playback speed, and therefore render times as well.

Systems:

Desktop

AMD Ryzen 7 1800x 8 core 16 thread at stock speed

64GB 3000mhz DDR4

Geforce RTX 3090

Windows 10

Laptop:

ASUS Zenbook Pro Duo 32GB (9980HK CPU, RTX 2060 GPU, dual 4K touch screens, main one OLED HDR)

Howard-Vigorita wrote on 8/10/2019, 12:39 AM

My only system with an Intel gpu onboard is the i9-9900k with AMD Radeon VIII. I tried it both ways both ways on that system and got identical timings. So I don't think they've actually implemented it yet in this version of Vegas 17. But I'm leaving it checked by default so I should see some changes in performance when they do implement it.

fr0sty wrote on 8/10/2019, 1:15 AM

AMD support for decode isn't in there yet, but quicksync is... but it is also reliant on the media being either avc or hevc, as that's all those chips can decode.

Systems:

Desktop

AMD Ryzen 7 1800x 8 core 16 thread at stock speed

64GB 3000mhz DDR4

Geforce RTX 3090

Windows 10

Laptop:

ASUS Zenbook Pro Duo 32GB (9980HK CPU, RTX 2060 GPU, dual 4K touch screens, main one OLED HDR)

Howard-Vigorita wrote on 8/10/2019, 2:09 AM

Ah, ha... the 4K Red Car uses 2 AVC 4k clips and I just tried it and the timings are indeed different with QSV decode disabled in file i/o.

FHD render w/QSV decode enabled: 51 sec; disabled: 1:11 sec.

But VP16 which does not have a File I/O settings section gets: 51 sec too. Looks like the File I/O setting for Intel QSV decoding is redundant in VP17. It's been a checkbox in General Settings in VP16 all along and now appears in 2 places in VP17. Looks like only the VP17 Nvidia capability is new. Fwiw, all my timings were run with both checked in VP17 and the one checked in VP16.

Howard-Vigorita wrote on 8/11/2019, 3:02 PM

AMD support for decode isn't in there yet, but quicksync is... but it is also reliant on the media being either avc or hevc, as that's all those chips can decode.

Just tested with output from ffmpeg prores_ks encoder which is essentially h.264/avc in a mov wrapper and no go on Vegas 16 or 17 using qsv encode/decode to speed things up. Exact same render times with qsv enabled or disabled. Here's what the MediaInfo looks like:

General
Complete name                            : D:\Red Car 4K\Media\Mercedes Clip 3 h264_422 2160p.mov
Format                                   : MPEG-4
Format profile                           : QuickTime
Codec ID                                 : qt   0000.02 (qt  )
File size                                : 3.07 GiB
Duration                                 : 4 min 30 s
Overall bit rate                         : 97.7 Mb/s
Writing application                      : Lavf58.27.103Video
ID                                       : 1
Format                                   : AVC
Format/Info                              : Advanced Video Codec
Format profile                           : High 4:2:2@L5.1
Format settings                          : 1 Ref Frames
Format settings, CABAC                   : No
Format settings, Reference frames        : 1 frame
Codec ID                                 : avc1
Codec ID/Info                            : Advanced Video Coding
Duration                                 : 4 min 30 s
Bit rate                                 : 96.6 Mb/s
Width                                    : 3 840 pixels
Height                                   : 2 160 pixels
Display aspect ratio                     : 16:9
Frame rate mode                          : Constant
Frame rate                               : 29.970 (30000/1001) FPS
Color space                              : YUV
Chroma subsampling                       : 4:2:2
Bit depth                                : 8 bits
Scan type                                : Progressive
Bits/(Pixel*Frame)                       : 0.388
Stream size                              : 3.04 GiB (99%)
Writing library                          : x264 core 157 r2970 5493be8
Encoding settings                        : cabac=0 / ref=1 / deblock=0:0:0 / analyse=0:0 / me=dia / subme=0 / psy=1 / psy_rd=1.00:0.00 / mixed_ref=0 / me_range=16 / chroma_me=1 / trellis=0 / 8x8dct=0 / cqm=0 / deadzone=21,11 / fast_pskip=1 / chroma_qp_offset=0 / threads=18 / lookahead_threads=3 / sliced_threads=0 / nr=0 / decimate=1 / interlaced=0 / bluray_compat=0 / constrained_intra=0 / bframes=0 / weightp=0 / keyint=250 / keyint_min=25 / scenecut=0 / intra_refresh=0 / rc=crf / mbtree=0 / crf=19.0 / qcomp=0.60 / qpmin=0 / qpmax=69 / qpstep=4 / ip_ratio=1.40 / aq=0
Language                                 : English
Codec configuration box                  : avcC

 

fr0sty wrote on 8/11/2019, 7:24 PM

I think QSV decode was actually present in Vegas 16 (I think it was automatic as long as the setting in general settings to use QSV when available was checked), 17 added support for NVDEC decoding as well on Nvidia Cards, and soon will add AMD support as well. That's why you notice similar times in 16.

I'm not sure what the decode limitations on the QSV and NVDEC chips are, it could be the 4:2:2 that is causing the issue. Worth reading up on.

Last changed by fr0sty on 8/11/2019, 7:25 PM, changed a total of 1 times.

Systems:

Desktop

AMD Ryzen 7 1800x 8 core 16 thread at stock speed

64GB 3000mhz DDR4

Geforce RTX 3090

Windows 10

Laptop:

ASUS Zenbook Pro Duo 32GB (9980HK CPU, RTX 2060 GPU, dual 4K touch screens, main one OLED HDR)

Howard-Vigorita wrote on 8/14/2019, 3:09 PM

Vegas 17 vs 16 benchmarks on the Intel Hades Canyon NUC (i7-8809g w/Radeon Vega/M & Intel HD630).

Note that in this case this is not an apples-to-apples comparison because when I did the v16 timings the NUC had Intel stock video drivers which lack the openCL api, making the Intel HD 630 iGpu invisible to Vegas 16 although it was still available to the cpu. Thus no QSV for rendering or decoding in Vegas. But last night I installed an older 2017 Intel driver from the online Windows Update archive that contains the openCL api and ran new timings with VP17. (as discussed in this thread). So the timing differences shown here between VP16 and VP17 are pretty much all attributable to the presence of QSV in VP17.

Also updated my Magic Table.

Howard-Vigorita wrote on 8/23/2019, 11:59 PM

I just got a Dell XPS-15 laptop. My 4 year-old Microsoft Surface Book bit the dust. Actually it kind of exploded. The Dell is better constructed (screwed, not glued) with an i7-8750h plus 32gb ram and a team of Intel UHD-630 plus Nvidia 1050-Ti internal gpus. Burned it in with some 4K Benchmarks using VP16 and VP17-Trial.

Was able to play the 4K Red Car proxy-project at a setting of Good-Auto at a solid 29.97 fps from beginning to end with both VP16 and 17... I have that rated 09 out of 16. Draft-Auto being an 01, Draft-Quarter 02, etc. See my Magic Table for the complete roundup of the various machines I've benched.

Note that although the Nvidia Studio drivers list the Nvidia 1050Ti as a supported board, they would not install on the Dell giving an error that the Windows version (Win10 v1903) is incompatible. Will update if that changes.