SVT-AV1 Encoder and Decoder | Hacker News Reader

SVT-AV1 Encoder and Decoder

gitlab.com

37 pointsskilled18 comments

If you need to do AV1 encoding then you should know the software encoder is dead slow.

Get a cheap Intel Arc GPU instead like the SPARKLE Intel Arc A310 or SPARKLE Intel Arc A380 ELF 6GB.

Almost all Hardwave video encoders come with the assumption of running in real time, thus sacrifice of quality for speed.

How much? Sometimes as much as switching to previous gen codec https://goughlui.com/2023/12/24/video-codec-round-up-2023-pa...

This is true for the AOM encoder, but SVT-AV1 is actually pretty darn fast in my experience!

libaom is slow because it's fundamentally a research product - the folks hacking on it are mostly concerned with the workload of encoding a basket of twenty short clips and demonstrating a tradeoff between Bjontegaard scores and increased encode time, and much less on holistically making a usable encoder.

Source: wrote part of the AV2 spec and traded off a 0.3% improvement in compression for a 1% increase in encode time

So where is AV2?

And 0.3% improvement for 1% increase in encode is pretty damn good.

True. From my testing SVT-AV1 is about 1.4x faster than x265, and about 540x faster than the original AOM encoder.

The benchmarks below show you need very powerful machines to get software encoding up to the sort of frame rates that hardware encoding will give you even on low end CPU/GPU combinations.

Even given the numbers in the benchmarks below I think the cheapest of Intel cards are going to make alot more sense than using the CPU of some honking great machine to do software av1 conversions.

As usual with benchmarks it takes some effort to understand what they mean.

Trying to get some tangible information into it with minimal effort:

Some benchmarks - use the dropdown to choose different preset/resolution combinations.

SVT-AV1 V2.1

Encoder Mode: Preset 13 - Input: Bosphorus 1080p https://openbenchmarking.org/test/pts/svt-av1&eval=0a8013353...

Encoder Mode: Preset 5 - Input: Bosphorus 1080p https://openbenchmarking.org/test/pts/svt-av1&eval=a618cfc70...

AOM AV1 3.9

Encoder Mode: Speed 11 Realtime - Input: Bosphorus 1080p https://openbenchmarking.org/test/pts/aom-av1&eval=70043aa1f...

Encoder Mode: Speed 6 Realtime - Input: Bosphorus 1080p https://openbenchmarking.org/test/pts/aom-av1&eval=fab03cd00...

Encoder Mode: Speed 6 Two-Pass - Input: Bosphorus 1080p https://openbenchmarking.org/test/pts/aom-av1&eval=fab03cd00...

I'm expecting next gen video codecs to all be neutral network based.

Ie. every frame is generated from the previous internal state and a few more bytes of input data which defines how the car moves or who the bad guy shoots.

Current state of the art systems require huge GPU's and minutes per frame. But smaller models will do a decentish job faster, and better hardware and algorithms are coming out every few months.

Sure, that's being predicted for about 10 years now. I think it's far more than one generation away simply because you'd need a decoder everywhere, and that takes a boatload of standardization. JpegXL failed to gain traction and got dumped for instance.

Best we can get next gen would be an experimental fork of x265 or some av1 codec with neural net baseline to improve on the defaut, like mkv was at first.

I worked at a startup (WaveOne) doing neural net-based compression. Your description of the state of the art seems a bit off, here's a 2021 paper [1] from WaveOne for comparison. Large GPUs, maybe, but only seconds (or less) for 720p+ frames.

[1] https://openaccess.thecvf.com/content/ICCV2021/papers/Rippel...

Will ‘lossy’ compression aspect act with accuracy of human memory - the attacker becomes a man, because it’s usually a man, etc?

Will the encoding of a movie give you different ethnicity for the bad guy?

We've basically already been there done that with jpeg2000 (edit: jbig2 not jpeg2000 as pointed out below)

You might be thinking of JBIG2

https://en.wikipedia.org/wiki/JBIG2#Character_substitution_e...

Reply on news.ycombinator.com