Discussion QSVEnc - HEVC

tectpro · Feb 14, 2023

I separated this part from the original conversation to not bloat that one unnecessarily.

I've got Intel's Arc A770, AMD and nVidia can suck on that

: (Intel's SDK scales up QP range with output bit depth, so in x265's terms QPs are 22/22/24)

Code:

QSVEncC (x64) 7.32 (r2990) by rigaya, Feb  7 2023 09:45:43 (gcc 9.4.0/Linux)
OS             Ubuntu 22.04.1 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 9 5950X 16-Core Processor (16C/32T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 22.4.3 (74f40ee)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.07
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x800, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 4 (high tier)
               1920x800p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 137208 frames, 377.51 fps, 893.36 kbps, 609.45 MB
encode time 0:06:03, CPULoad: 0.0%
frame type IDR    572
frame type I     1144,  total size   43.59 MB
frame type P     8575,  total size  184.09 MB
frame type B   128061,  total size  403.56 MB

0x0x0x0x0 said:
Code:

~/w/ffmpeg/ffmpeg -i ~/w/src.h264 -vf crop=1920:800:0:140 -strict -1 -pix_fmt yuv420p10le -f yuv4mpegpipe -v 0 - | ~/w/qsvencc/qsvencc --y4m -i - -c hevc -u best --profile main10 --cqp 34:34:36 --i-adapt --b-adapt --scenario-info archive --open-gop -b 16 --output-depth 10 --la-depth 100 --la-quality slow --pic-struct --hevc-gpb --async-depth 8 -o encoded.hevc

Bear in mind, on a 5950X ffmpeg piping to qsvencc is faster than qsvencc doing decoding and encoding, on other processors it might be slower to way slower.

Tested it with the A770
This is with pipe from ffmpeg - no hyper mode enabled in QSVenc

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [5.00GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 131 work buffer
Input Info     avsw: rawvideo(yv12)->p010 [AVX2], 1920x1080, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
frame=34047 fps=253 q=-0.0 Lsize=103442308kB time=00:23:40.08 bitrate=596724.0kbits/s speed=10.6x
video:103417762kB audio:22189kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.002277%
encoded 34047 frames, 257.70 fps, 1144.54 kbps, 193.75 MB
encode time 0:02:12, CPU: 0.5%, GPU: 23.1%, VD: 99.1%
frame type IDR   142
frame type I     142,  total size   11.96 MB
frame type P    2128,  total size   69.73 MB
frame type B   31777,  total size  112.06 MB

This is normal QSVEnc does both jobs - no hyper mode enabled in QSVenc

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [5.02GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 160 work buffer
Input Info     avqsv: H.264/AVC, 1920x1080, 24000/1001 fps
VPP            ColorFmtConvertion: nv12 -> p010
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
encoded 34047 frames, 268.98 fps, 1158.52 kbps, 196.12 MB
encode time 0:02:06, CPU: 0.3%, GPU: 30.8%, VD: 117.2%
frame type IDR   142
frame type I     142,  total size   11.96 MB
frame type P    2128,  total size   69.82 MB
frame type B   31777,  total size  114.34 MB

Hyper mode doesn't seem to work with HEVC.

Retesting with HDD 1 to HDD 2 and SSD to SSD
This test was HHD 1 to HDD1

The 8 cores more in the Ryzen 9 5950X make a difference.

tectpro · Feb 14, 2023

2nd test HDD 1 to HDD 2
This is with pipe from ffmpeg - no hyper mode enabled in QSVenc

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [5.03GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 131 work buffer
Input Info     avsw: rawvideo(yv12)->p010 [AVX2], 1920x1080, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
frame=34047 fps=248 q=-0.0 Lsize=103442308kB time=00:23:40.08 bitrate=596724.0kbits/s speed=10.3x
video:103417762kB audio:22189kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.002277%
encoded 34047 frames, 252.39 fps, 1144.54 kbps, 193.75 MB
encode time 0:02:14, CPU: 0.5%, GPU: 24.9%, VD: 98.8%
frame type IDR   142
frame type I     142,  total size   11.96 MB
frame type P    2128,  total size   69.73 MB
frame type B   31777,  total size  112.06 MB

This is normal QSVEnc does both jobs - no hyper mode enabled in QSVenc

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [4.92GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 160 work buffer
Input Info     avqsv: H.264/AVC, 1920x1080, 24000/1001 fps
VPP            ColorFmtConvertion: nv12 -> p010
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
encoded 34047 frames, 263.65 fps, 1158.52 kbps, 196.12 MB
encode time 0:02:09, CPU: 0.3%, GPU: 30.0%, VD: 115.2%
frame type IDR   142
frame type I     142,  total size   11.96 MB
frame type P    2128,  total size   69.82 MB
frame type B   31777,  total size  114.34 MB

No test done for SSD to SSD yet.
To be done later.

tectpro · Feb 14, 2023

@0x0x0x0x0
The quality looks good

Even additionally adding resizing 720p to 1080p and convolution3d does not slow down it seems.

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [4.92GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 169 work buffer
Input Info     avqsv: H.264/AVC, 1280x720, 24000/1001 fps
VPP            cspconv(nv12 -> yv12(16bit))
               convolution3d: matrix standard, mode normal
                       threshold spatial luma 3, chroma 4, temporal luma 3, chroma 4
               cspconv(yv12(16bit) -> p010)
               Resize 1280x720 -> 1920x1080
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
encoded 32394 frames, 262.73 fps, 1269.10 kbps, 204.41 MB
encode time 0:02:03, CPU: 0.3%, GPU: 25.6%, VD: 106.5%
frame type IDR   135
frame type I     135,  total size    8.09 MB
frame type P    2024,  total size   54.56 MB
frame type B   30235,  total size  141.75 MB

Seems great for cartoons, animations, animes

tectpro · Feb 14, 2023

And one more upscale from 720p to 2160p - still impressive performance

Code:

QSVEncC (x64) 7.34 (r2999) by rigaya, Feb 13 2023 12:32:42 (VC 1934/Win)
OS             Windows 11 x64 (22621) [UTF-8]
CPU Info       12th Gen Intel Core i9-12900K [4.89GHz] (8P+8E,16C/24T) <DG2>
GPU Info       Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4091)
Media SDK      QuickSyncVideo (hardware encoder) FF, 2nd GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  d3d11, 150 work buffer
Input Info     avqsv: H.264/AVC, 1280x720, 24000/1001 fps
VPP            cspconv(nv12 -> yv12(16bit))
               convolution3d: matrix standard, mode normal
                       threshold spatial luma 3, chroma 4, temporal luma 3, chroma 4
               resize(spline36): 1280x720 -> 3840x2160
               cspconv(yv12(16bit) -> p010)
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 6 (high tier)
               3840x2160p 1:1 23.976fps (24000/1001fps)
               avwriter: hevc, aac => matroska
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all
encoded 35890 frames, 71.49 fps, 2856.06 kbps, 509.65 MB
encode time 0:08:22, CPU: 1.2%, GPU: 22.5%, VD: 102.2%
frame type IDR   150
frame type I     150,  total size   15.43 MB
frame type P    2243,  total size  124.02 MB
frame type B   33497,  total size  370.20 MB

ssim/psnr: SSIM YUV: 0.996260 (24.271188), 0.998130 (27.281482), 0.998103 (27.218622), All: 0.996879 (25.056672), (Frames: 35890)

ssim/psnr: PSNR YUV: 51.352604, 54.621474, 54.466544, Avg: 52.179905, (Frames: 35890)

cartman0208 · Feb 14, 2023

Meh, I just got a 3070ti ... do I really need to get an additional ARC770 (even if it's half the price) to have the best speed (and quality) for the buck?

tectpro · Feb 15, 2023

cartman0208 said:
Meh, I just got a 3070ti ... do I really need to get an additional ARC770 (even if it's half the price) to have the best speed (and quality) for the buck?

I don't think that, that's needed.
The quality of the rtx is not bad and the speed is still fast.
I will try post one of my cmd lines for Nvidia NVEnc and HEVC when I got time.
Than you can test with that.

If you are interested.
Currently Intel has a promotion on top of their price reduction.

Code:

[LIST]
[*]Intel 23Q1-02
[/LIST]

Highlights
[LIST]
[*]Get 2 Games & 5 Creative Apps
[*]Start Date: February 1, 2023
[*]End Date: April 30, 2023
[*]Redemption End Date: July 31, 2023
[/LIST]
Get two full games and five creative apps when purchasing a qualifying Intel Arc 3, 5, 7 AIC/IBC/DT, or MB system with the Intel Arc Spring Play & Create Bundle.


This bundle is available starting February 1, 2023 and ends on April 30, 2023. If you receive a key in the eligible timeframe, you have until July 31, 2023 to redeem it. For more details on how to redeem, please visit the official Intel website.

This bundle includes the following:

[LIST]
[*]Nightingale (full game, RP - ESRB Rating)
[*]The Settlers: New Allies (full game, T - ESRB Rating)
[*]PowerDirector 365 (1-year subscription)
[*]XSplit Premium (3-month subscription)
[*]Topaz Gigapixel AI (full download)
[*]D5 Render (3-month subscription)
[*]MAGIX Video Pro X14 (6-month subscription)
[/LIST]

cartman0208 · Feb 15, 2023

Currently I'm using something like this to resize to 720p and have a decent quality with 2-2.5 mbits video bitrate. I get around 320 fps during the process.

Code:

--avhw  --cqp 22 --codec h265 --preset P7 --output-depth 10 --qp-init 22 --qp-max 24 --qp-min 20 --max-bitrate 0 --aq --aq-temporal --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 4 --multiref-l1 4 --strict-gop --nonrefp --weightp --vpp-deband  --vpp-pmd  --vpp-warpsharp  --output-res 1280x720

0x0x0x0x0 · Feb 18, 2023

cartman0208 said:
Meh, I just got a 3070ti ... do I really need to get an additional ARC770 (even if it's half the price) to have the best speed (and quality) for the buck?

I actually got the Arc for AV1, but HEVC was way too got to not abuse it

0x0x0x0x0 · Mar 11, 2023

tectpro said:
[snip]

The 8 cores more in the Ryzen 9 5950X make a difference.

Now that I got my hands on a 5800x3d with Amazon having a fire-sale on them, I swapped 5950x with 5800x3d, the rest of the system remained the same:-

5950x:-

Code:

QSVEncC (x64) 7.32 (r2990) by rigaya, Feb  7 2023 09:45:43 (gcc 9.4.0/Linux)
OS             Ubuntu 22.04.2 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 9 5950X 16-Core Processor (16C/32T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 23.1.0 (012af72)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x800, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 4 (high tier)
               1920x800p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 137208 frames, 378.86 fps, 893.36 kbps, 609.45 MB
encode time 0:06:02, CPULoad: 0.0%
frame type IDR    572
frame type I     1144,  total size   43.59 MB
frame type P     8575,  total size  184.09 MB
frame type B   128061,  total size  403.56 MB

5800x3d:-

Code:

QSVEncC (x64) 7.32 (r2990) by rigaya, Feb  7 2023 09:45:43 (gcc 9.4.0/Linux)
OS             Ubuntu 22.04.2 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 7 5800X3D 8-Core Processor (8C/16T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 23.1.0 (012af72)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.08
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x800, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 4 (high tier)
               1920x800p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  WeightP WeightB GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 137208 frames, 381.12 fps, 893.36 kbps, 609.45 MB
encode time 0:06:00, CPULoad: 0.0%
frame type IDR    572
frame type I     1144,  total size   43.59 MB
frame type P     8575,  total size  184.09 MB
frame type B   128061,  total size  403.56 MB

Totally unscientific, but make of that what you will... both tests were ran on "performance" scaling governor.

cartman0208 · Mar 11, 2023

Despite lesser cores it's faster?

Higher base clock?

0x0x0x0x0 · Mar 11, 2023

+64MiB L3 cache advantage on the chiplet (total L3 cache is 96MiB): 5800x3d has a single chiplet of 8C w/ 96MiB L3 cache, 5950x has two chiplets of 8C each with 32MiB L3 cache on each. Both have base clock at 3.4GHz, and 5950x has a +400MHz boost advantage.

RedFox 1 · Mar 12, 2023

tectpro said:
QSVEncC (x64) 7.34 (r2999) by rigaya

The latest QSVEncC is 7.35, have you updated and tried that yet?

0x0x0x0x0 · Mar 12, 2023

RedFox 1 said:
The latest QSVEncC is 7.35, have you updated and tried that yet?

Code:

QSVEncC (x64) 7.35 (r3018) by rigaya, Mar  7 2023 12:48:16 (gcc 11.3.0/Linux)
OS             Ubuntu 22.04.2 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 7 5800X3D 8-Core Processor (8C/16T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 23.1.0 (012af72)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.07
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x800, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 4 (high tier)
               1920x800p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 137208 frames, 380.99 fps, 886.38 kbps, 604.69 MB
encode time 0:06:00, CPULoad: 0.0%
frame type IDR     10
frame type I      582,  total size   21.94 MB
frame type P     8013,  total size  177.51 MB
frame type B   128623,  total size  405.45 MB

RedFox 1 · Mar 12, 2023

Very impressive.

Ch3vr0n · Mar 12, 2023

380fps? damn lol

0x0x0x0x0 · Mar 12, 2023

Ch3vr0n said:
380fps? damn lol

Yup, clean 10 bit input (2022 film) with upto 16 concurrent b-frames and pyramid, but that's on a cropped frame (1920 x 800), on full frame (much noisier 1995 film) you get:-

Code:

QSVEncC (x64) 7.35 (r3018) by rigaya, Mar  7 2023 12:48:16 (gcc 11.3.0/Linux)
OS             Ubuntu 22.04.2 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 7 5800X3D 8-Core Processor (8C/16T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 23.1.0 (012af72)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.07
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x1080, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 184434 frames, 142.78 fps, 2055.98 kbps, 1885.35 MB
encode time 0:21:31, CPULoad: 0.0%
frame type IDR     13
frame type I      782,  total size    54.56 MB
frame type P    10771,  total size   488.89 MB
frame type B   172894,  total size  1342.70 MB

I like my b-frames, which is why I was prepared to tolerate x265 over NVENC, and then you throw in Intel's adaptive LTR... no brainer!

RedFox 1 · Mar 12, 2023

0x0x0x0x0 said:
encoded 137208 frames, 381.12 fps, 893.36 kbps, 609.45 MB

Unreal, I never saw that in my life.

0x0x0x0x0 · Mar 12, 2023

RedFox 1 said:
Unreal, I never saw that in my life.

It's rather good, my Arcs have been getting a good amount of abuse, and it seems to be hardly any difference between Intel's own A770 Limited Edition and Asrock's A770 (overclocked plain A770). I've not tried Acer's one, it's still stupidly expensive compared to the other two, and Intel's supposedly bringing Alchemist refresh in Q3 and Battlemage in '24Q1... But for complete "picture," here's clean modern animation (full frame):-

Code:

QSVEncC (x64) 7.35 (r3018) by rigaya, Mar  7 2023 12:48:16 (gcc 11.3.0/Linux)
OS             Ubuntu 22.04.2 LTS (5.17.0-1026-oem) x64
CPU Info       AMD Ryzen 7 5800X3D 8-Core Processor (8C/16T) <DG2>
GPU Info       Intel Graphics / Driver : Intel iHD driver for Intel(R) Gen Graphics - 23.1.0 (012af72)
Media SDK      QuickSyncVideo (hardware encoder) FF, 1st GPU, API v2.07
Async Depth    8 frames
Hyper Mode     off
Buffer Memory  va, 131 work buffer
Input Info     y4m(yv12(10bit))->p010 [AVX2], 1920x1080, 24000/1001 fps
AVSync         cfr
Output         HEVC(yuv420 10bit) main10 @ Level 5 (high tier)
               1920x1080p 1:1 23.976fps (24000/1001fps)
Target usage   1 - best
Encode Mode    Constant QP (CQP)
CQP Value      I:34  P:34  B:36
Scenario Info  archive
QP Limit       min: 22, max: 63
Ref frames     6 frames
Bframes        16 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features  GPB QPOffset pic_struct tskip ctu:64 sao:all

encoded 31684 frames, 203.53 fps, 1277.47 kbps, 201.25 MB
encode time 0:02:35, CPULoad: 0.0%
frame type IDR     3
frame type I     136,  total size   13.86 MB
frame type P    1851,  total size   68.88 MB
frame type B   29700,  total size  118.66 MB

lol:-

Code:

-r--r--r-- 1 xxx  xxx   211020705 Mar 12 23:32 BD_X.T1.hevc
-r--r--r-- 1 xxx  xxx    99475460 Mar 13 00:00 BD_X.T1.m4a

audio from

Code:

qaac 2.79, CoreAudioToolbox 7.10.9.0

BD_X.T1.m4a
AAC-LC Encoder, TVBR q127, Quality 96
[100.0%] 22:01.453/22:01.453 (13.7x), ETA 0:00.000
63429766/63429766 samples processed in 1:36.438
Overall bitrate: 600.65kbps

cartman0208 · May 23, 2023

Did anyone play around with encoding settings for AV1?
I'm currently at CQP 40 and the resulting file is just marginally smaller than the source.
That differs greatly from HEVC, where I could see immense size differences already from 20 to 24

Code:

.\nvencc64 --avhw  --cqp 40 --codec av1 --preset P7 --output-depth 10 --profile main10 --qp-init 40 --qp-max 50 --qp-min 30 --aq --aq-temporal --bframes 5 --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 3 --multiref-l1 3 --strict-gop --nonrefp --vpp-deband  --vpp-pmd  --vpp-warpsharp

BUT the speed is surprisingly better ... I get around 50fps for a 4K output in HEVC, for AV1 that's 90fps

tectpro · May 23, 2023

cartman0208 said:
Did anyone play around with encoding settings for AV1?
I'm currently at CQP 40 and the resulting file is just marginally smaller than the source.
That differs greatly from HEVC, where I could see immense size differences already from 20 to 24

Code:

.\nvencc64 --avhw --cqp 40 --codec av1 --preset P7 --output-depth 10 --profile main10 --qp-init 40 --qp-max 50 --qp-min 30 --aq --aq-temporal --bframes 5 --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 3 --multiref-l1 3 --strict-gop --nonrefp --vpp-deband --vpp-pmd --vpp-warpsharp

BUT the speed is surprisingly better ... I get around 50fps for a 4K output in HEVC, for AV1 that's 90fps

Try VBR instead of CQP for AV1.
See if that makes a difference.
I can't test it for NVENC.
There's still only 5 bframes max for nvidia?

Why are you using slices and strict-gop?
Lookahead 16 is that the max for AV1? Isnt it 32? And the multi ref is it limited to 3?

Please see my questioning explained below:

Using slices has a few disadvantages:

Quality Loss: Each slice is independently encoded and decoded. Therefore, it's not possible to predict one slice from another, which limits the efficiency of the inter prediction and could lead to a decrease in compression efficiency.

Complexity: The use of slices increases the complexity of the encoding and decoding processes, and may require more computational resources.

Strict GOP may have slightly worse compression efficiency because it can't take advantage of temporal redundancy across GOP boundaries.

A lookahead value of 32 means that the encoder will analyze the next 32 frames before encoding each frame. This can potentially lead to better compression efficiency because it gives the encoder more information about upcoming scenes, allowing for better bit allocation and frame type decision-making. However, it will also increase computational demand, as more frames need to be analyzed in advance.

Also for instance with --multiref-l0 7 --multiref-l1 7, would potentially improve compression efficiency. The encoder would have more options for prediction, allowing it to more accurately encode changes between frames.

However, the trade-off is increased computational complexity and potentially longer encoding times.

Discussion QSVEnc - HEVC

Translator (ms_MY)

Translator (ms_MY)

Translator (ms_MY)

Translator (ms_MY)

Moderator

Translator (ms_MY)

Moderator

Well-Known Member

Well-Known Member

Moderator

Well-Known Member

Forum Admin

Well-Known Member

Forum Admin

Translator NL

Well-Known Member

Forum Admin

Well-Known Member

Moderator

Translator (ms_MY)