Xaymar
patreon
Unfiltered preliminary results from Encoder Showdown 2024 - Round 1 (Gaming Footage)
🕑 Added 2024-08-08 03:02:29 +0000 UTCThe first pass for Encoder Showdown 2024 has finished. I've excluded software encoders, as software encoders for modern formats already outclass all GPU encoders by default. These are the configurations that scored the highest on 90% of encoded frames:
AMD AMF
AMF received a score of 66.38% using VMAF 4k, and a score of 58.28% using VMAF 1080p. Note: AMD provides two ways to engage the high quality mode of their encoder, hqcbr and usage/quality.
Overall: -preencode true -enforce_hrd true -filler_data false -preanalysis true -pa_activity_type yuv -pa_scene_change_detection_enable true -pa_scene_change_detection_sensitivity high -pa_static_scene_detection_enable true -pa_static_scene_detection_sensitivity medium -pa_caq_strength high -pa_frame_sad_enable true -pa_ltr_enable true -pa_lookahead_buffer_depth 29 -pa_paq_mode caq -pa_taq_mode 1 -pa_high_motion_quality_boost_mode auto -pa_max_qp_before_force_skip 51 -bf 3
AV1: -usage transcoding -profile main -level auto -quality high_quality -rc cbr -header_insertion_mode gop -high_motion_quality_boost_enable false -skip_frame false -align 1080p
H265/HEVC: -usage high_quality -profile main -profile_tier high -level auto -quality quality -rc cbr -header_insertion_mode idr -vbaq true -skip_frame false -me_half_pel true -me_quarter_pel true -high_motion_quality_boost_enable false
H264/AVC: -usage high_quality -profile high -level auto -quality quality -rc cbr -frame_skipping false -vbaq true -me_half_pel true -me_quarter_pel true -coder cabac -bf_ref true
Notable Things:
high_motion_quality_boost_enable and pa_high_motion_quality_boost_mode appear to be polar opposites. The former makes the video quality significantly worse, the latter improves it slightly.
vbaq slightly reduces overall quality, but improves it in areas where its really needed.
Intel QSV
QSV achieved a score of 78.88% with VMAF 4k, and a score of 82.42% using VMAF 1080p. QSV is the only encoder - aside from x264 - where dropping to a faster VMAF model results in an increased score.
Overall: -preset veryslow -look_ahead_depth 100 -bf 3
AV1: -adaptive_i 1 -adaptive_b 1 -b_strategy 1 -extbrc 1 -low_delay_brc 0 -profile main
H265/HEVC: -rdo 1 -mbbrc 1 -scenario gamestreaming -skip_frame no_skip -dual_gfx off -profile main -tier high -gpb true -adaptive_i true -adaptive_b true
H264/AVC: -rdo 1 -mbbrc 1 -scenario gamestreaming -skip_frame no_skip -dual_gfx off -cavlc false -vcm false -look_ahead true -look_ahead_downsampling auto -profile high
VP9: Good joke, this just crashes all the time. Does it even exist in hardware?
NVIDIA NVENC
NVENC achieved a score of 92.44% using VMAF 4k, and a score of 92.28% using VMAF 1080p.
Overall: -preset p7 -tune hq -rc cbr -no-scenecut false -b_adapt true -spatial_aq false -temporal_aq true -strict_gop false -bf 3 -rc-lookahead 30 -zerolatency false -nonref_p false -init_qpP 0 -init_qpB 0 -init_qpI 0 -b_ref_mode each -multipass fullress -lookahead_level 3
AV1: -tier 1-strict_gop true -split_encode_mode disabled -highbitdepth true
H265/HEVC: -profile main -tier high -strict_gop false -split_encode_mode disabled -tf_level 4 -highbitdepth true -unidir_b true
H264/AVC: -profile high -strict_gop false
Other Things
My tests rely on FFmpeg, so mileage may vary, but so far it's pretty much a carbon copy of the previous years. AMD is trailing behind significantly, QSV enjoys being right in the middle, and NVENC is leading with a massive distance. Software Encoders (in real-time configurations) tend to hover well behind NVENC at this point, usually being equal to QSV.
H264: NVENC > x264 placebo, veryslow, slower > QSV > x264 slow, medium, fast, faster > AMF > x264 veryfast, superfast, ultrafast
H265: NVENC > QSV > AMF
AV1: SVT-AV1 preset 8 and lower > NVENC > SVT-AV1 preset 9, 10 > QSV > SVT-AV1 preset 11 > AMF > SVT-AV1 preset 12
I'm not quite sure why AMD AMF has such trouble with AV1, but at least they've finally surpassed x264 veryfast. Though the encode did take 5 times as long, and was just barely compatible with cinematic framerates (25/30 fps).
I'll still need to run more rounds, particularly ones with overlaid streamer layouts and other moving parts, and then have to go through results and manually inspect footage and results - VMAF tends to be biased towards software encoder artifacting, instead of true quality.
Comments
Vezart of Vhonowslend
https://www.xaymar.com/articles/2022/01/10/h264-encoder-showdown/ I used to test and compare encoders in quality metrics. The above are the settings that scored the best quality in more than 90% of encoded frames.
Tsumi
I'm missing the context, what is this all? it seems interesting
Vezart of Vhonowslend
For those wishing to know just how much video data I'll need to filter through, I have a NAS that can hold 20 TiB of data. When I started the project, it had 18.2 TiB free. It now has 44 GiB free, after having to delete the footage twice already so it could continue. NVENC and AMF adding more options has ballooned the necessary testing area significantly. If I had any background in machine learning, I might be able to use one of those gradient descent thingies to optimize this.
