Intel SVT-HEVC encoder test

[sophisticles]

Hello guys,

It's been a long time since I posted an encoder test and I decided to break the drought with a test of Intel's new open Source SVT-HEVC encoder, which is part of a family of encoders that also includes a VP9 and an AV1 encoder.

Some of you may have seen me mention these encoders in other threads, their main claim to fame is that they are built from the ground up with threading and AVX-512 in mind.

Phoronix did a couple of tests of these encoders and the performance was astounding:

https://www.phoronix.com/scan.php?page=news_item&px=SVT-AV1-Xeon-EPYC

https://www.phoronix.com/scan.php?page=news_item&px=27-CPUs-AV1-SVT-Video-Encode

Using the SVT-HEVC encoder a 7900X is capable of achieving 200fps 1080p encode, while a 2990WX, the fastest of all the processor in the x264 test, only achieved 146fps in that type of test.

For this test, I went to https://media.xiph.org/video/derf/ and downloaded every 1080p source file and created 3 different source files, a 1080p, a 1080p25 and a 1080p50 file by important the appropriate files into Shotcut and exporting using Huffy.

I downloaded the latest ffmpeg snapshot, patched it so that it could use Intel's SVT-HEVC encoder and then conducted 3 tests for each source. The first was a baseline using x264 with the default settings, with the following command lines:

time ./ffmpeg -i 1080p_Source.mkv -c:v libx264 Test_Encode_1080_x264.mkv, 2m6.862s
time ./ffmpeg -i 1080p25_Source.mkv -c:v libx264 Test_Encode_1080p25_x264.mkv, 1m14.061s
time ./ffmpeg -i 1080p50_Source.mkv -c:v libx264 Test_Encode_1080p50_x264.mkv, 1m9.294s

This gave me a baseline with regard to both bitrate used and encoding speed. I then chose settings for x265 and SVT-HEVC that matched both the encoding speed and bitrate used by x264. The resulting files are posted here.

A few notes:

My system, Ryzen 5 1600 + 8GB DDR4-2400 does not support AVX-512, with a processor that did the results of this test would greatly swing in the favor of the Intel encoder. The reason for this is that x264 sees about a 20% performance gain from AVX-512, x265 sees about a 30% uplift but the SVT-HEVC encoder sees a monstrous increase from AVX-512 (compare the 9900K vs 7900X). Because of this performance gain, using the testing parameters I used, using an AVX-512 enabled would allow me to use much higher quality settings for the Intel encoder compared to x264/x265 and still realize much faster performance.

I also plan on uploading one encode of each test file with the highest quality settings available for the Intel encoder.

time ./ffmpeg -i 1080p_Source.mkv -c:v libx264 Test_Encode_1080_x264.mkv
2m6.862s

time ./ffmpeg -i 1080p_Source.mkv -c:v libx265 -b:v 10.7M -preset ultrafast Test_Encode_1080_x265.mkv
1m51.399s

./ffmpeg -i 1080p_Source.mkv -c:v libsvt_hevc -rc 1 -b:v 10.7M -preset 7 -tune 0 -la_depth 40 Test_Encode_1080_SVT-HEVC.mkv
1m57.054s

time ./ffmpeg -i 1080p25_Source.mkv -c:v libx264 Test_Encode_1080p25_x264.mkv
1m14.061s

time ./ffmpeg -i 1080p25_Source.mkv -c:v libx265 -b:v 7.3M -preset ultrafast Test_Encode_1080p25_x265.mkv
1m13.026s

time ./ffmpeg -i 1080p25_Source.mkv -c:v libsvt_hevc -rc 1 -b:v 7.3M -preset 7 -tune 0 -la_depth 40 Test_Encode_1080p_25_SVT-HEVC.mkv
0m43.913s

time ./ffmpeg -i 1080p50_Source.mkv -c:v libx264 Test_Encode_1080p50_x264.mkv
1m9.294s

time ./ffmpeg -i 1080p50_Source.mkv -c:v libx265 -b:v 22.7M -preset ultrafast Test_Encode_1080p50_x265.mkv
1m16.305s

time ./ffmpeg -i 1080p50_Source.mkv -c:v libsvt_hevc -rc 1 -b:v 22.7M -preset 7 -tune 0 -la_depth 40 Test_Encode_1080p50_SVT-HEVC.mkv
1m11.967s

The Intel encoder supports 3 different tune settings, tune 0 is optimized for subjective visual quality, tune 1 is optimized for PSNR/SSIM and tune 2 is optimized for VMAF, the SVT-HEVC supports AQ but only for 4k or higher resolutions. and the presets range from 0 to 12 or 0 to 10 depending on the tune and resolution.

I'll let you guys decide for yourselves how SVT-HEVC compares.