Question Intel SVT encoders

[Carfax83]

I'm not too much into encoding/transcoding, but I just came across a new line of codecs Intel is developing called SVT, or scalable video technology. Anyone that's ever done encoding/transcoding through software knows that the process can be long and laborious dependent on the size of the source file and the quality of the output, but despite several hardware solutions that are exceedingly fast, many still prefer to go through the software route for higher quality.

But what if software encoding/transcoding could be done many times faster while still achieving high quality? That's what Intel's SVT hopes to accomplish. As for the performance, I was literally stunned. Intel seems to be leveraging multithread and SIMD optimization to a very high degree to be achieving such framerates. Here are some benchmarks that Phoronix has done to illustrate the massive performance boost versus competing methods:

27 CPUs Benchmarked With AOM AV1, Intel SVT VP9/AV1/HEVC Video Encoders
The Intel SVT-VP9 Performance Boost Across 10 Intel/AMD Systems

I find it fascinating that the CPUs that lack AVX2 in the benchmarks are for want of a better term, simply obliterated! And even AVX-512 makes an excellent showing. In the second benchmark, the Intel 7980Xe is able to be extremely competitive with the Zen 2 Epyc CPUs despite having far less cores, and this is primarily due to AVX-512.

Now my next question is, has anyone ever tried these codecs to see if the quality is comparable to other branches? It must be up to par because Netflix has partnered with Intel and will be using their SVT-AV1 encoder.

 

[soresu]

I'm not too much into encoding/transcoding, but I just came across a new line of codecs Intel is developing called SVT, or scalable video technology. Anyone that's ever done encoding/transcoding through software knows that the process can be long and laborious dependent on the size of the source file and the quality of the output, but despite several hardware solutions that are exceedingly fast, many still prefer to go through the software route for higher quality.

But what if software encoding/transcoding could be done many times faster while still achieving high quality? That's what Intel's SVT hopes to accomplish. As for the performance, I was literally stunned. Intel seems to be leveraging multithread and SIMD optimization to a very high degree to be achieving such framerates. Here are some benchmarks that Phoronix has done to illustrate the massive performance boost versus competing methods:

27 CPUs Benchmarked With AOM AV1, Intel SVT VP9/AV1/HEVC Video Encoders
The Intel SVT-VP9 Performance Boost Across 10 Intel/AMD Systems

I find it fascinating that the CPUs that lack AVX2 in the benchmarks are for want of a better term, simply obliterated! And even AVX-512 makes an excellent showing. In the second benchmark, the Intel 7980Xe is able to be extremely competitive with the Zen 2 Epyc CPUs despite having far less cores, and this is primarily due to AVX-512.

Now my next question is, has anyone ever tried these codecs to see if the quality is comparable to other branches? It must be up to par because Netflix has partnered with Intel and will be using their SVT-AV1 encoder.

It's not on par with libaom for quality as yet, but much, MUCH faster - especially with the new Epyc Rome 7742 CPU.

Netflix partnered with SVT-AV1 because of the speed, the quality can likely come faster for that than speed will for libaom - libaom is more like a development platform for a codec that had lots of optimisations added, whereas SVT-AV1 was designed for production encoding.

As for non-AVX2 CPU's getting obliterated, it likely doesn't have much non-AVX SIMD code, ie SSSE3, or SSE4, so anything before AVX2 would simply be running at scalar code paths if I'm right.

As to 7980XE doing so well, I get the impression that SVT-AV1 does have a parallel scaling ceiling/bottleneck somewhere, because 2x7742 are only a few fps faster than 1x7742 in the last comparison with them I read.

 

[Carfax83]

It's not on par with libaom for quality as yet, but much, MUCH faster - especially with the new Epyc Rome 7742 CPU.

Yeah AMD Epyc CPUs that use the Zen 2 core got a nice boost with that update. The Epyc 7551 however had a very minor performance boost, due to more anemic SIMD capabilities.

Netflix partnered with SVT-AV1 because of the speed, the quality can likely come faster for that than speed will for libaom - libaom is more like a development platform for a codec that had lots of optimisations added, whereas SVT-AV1 was designed for production encoding.

I wonder how on Earth is Intel getting such large performance increases in leaps and bounds, while the competitors seem to be stagnant? Was there really this much untapped performance in the code for these codecs?

As for non-AVX2 CPU's getting obliterated, it likely doesn't have much non-AVX SIMD code, ie SSSE3, or SSE4, so anything before AVX2 would simply be running at scalar code paths if I'm right.

But Sandy Bridge and Ivy Bridge support AVX, as does the FX-8370e. I suppose the codecs don't use any floating point though, as AVX is floating point only if I'm not mistaken.

As to 7980XE doing so well, I get the impression that SVT-AV1 does have a parallel scaling ceiling/bottleneck somewhere, because 2x7742 are only a few fps faster than 1x7742 in the last comparison with them I read.

Yeah I agree about the parallel scaling bottleneck, but I was talking primarily about the SVT-VP9 codec and not the AV1. VP9 seems to respond particularly well to wider vectors.

 

[soresu]

I wonder how on Earth is Intel getting such large performance increases in leaps and bounds, while the competitors seem to be stagnant? Was there really this much untapped performance in the code for these codecs?

SVT is a company Intel bought, they seem to have some sort of better trick to parallel scaling, as whatever performance bottlenecks it has, it doesn't seem to lose quality as it scales to more threads.

A recent conference called Big Apple Video demonstrated the quality loss per extra thread for several codecs, some do better than others but SVT just stays straight from the start, the guy implied he was intent on working it out from the code and sticking it in his own.

I believe the quality loss is a significant reason that others have not pursued more threads further.

 

[soresu]

Something to note as well, the feature of keeping quality as threads/speed scales would be a killer feature to a company like Netflix.

Some companies may not seem to care much about their encoding output quality, but from what I have gathered Netflix are extremely on the ball here.

The only other company I've noted such a dedication from are the BBC and their iPlayer platform, which is great quality wise, and has consistently outperformed competitors in the UK.

 

[Carfax83]

SVT is a company Intel bought, they seem to have some sort of better trick to parallel scaling, as whatever performance bottlenecks it has, it doesn't seem to lose quality as it scales to more threads.

A recent conference called Big Apple Video demonstrated the quality loss per extra thread for several codecs, some do better than others but SVT just stays straight from the start, the guy implied he was intent on working it out from the code and sticking it in his own.

I believe the quality loss is a significant reason that others have not pursued more threads further.

Yeah I've been reading on their website about how they accomplish it. It's pretty amazing if you ask me. Do you think it could be the death knell of fixed function hardware encoders? I remember years ago, that people were saying that GPUs were going to replace CPUs for encoding/transcoding. What an ironic turn of events!

As for multiple threads leading to loss of quality, what about the SIMD optimization? That seems to be the biggest part of their performance increase, judging from the benchmarks at least.

 

[soresu]

I remember years ago, that people were saying that GPUs were going to replace CPUs for encoding/transcoding. What an ironic turn of events!

I remember getting my first full PC build in 2003, then in a mere 3 years I had an Athlon X2, and a Phenom II X4 in 2009, even so I didn't expect 64 cores would ever happen back then (evidently neither did Intel management!).

 

[soresu]

As for multiple threads leading to loss of quality, what about the SIMD optimization? That seems to be the biggest part of their performance increase, judging from the benchmarks at least.

libaom has a lot o SIMD too, I don't think it's that alone - the sustainable scaling allows the encoder to make full use of all that SIMD power without sacrificing the stream quality.

There may be some tricks in there to better assign more threads too, possibly segment/chunk encoding - that might gel with the memory increases as you scale cores, though I probably just got the wrong end of that stick.

 

[beginner99]

While better speed at equal quality obviously makes huge sense for netflix, still the streaming quality is usually pretty bad. I mean youtube 1080p can easily look much worse than a good 720p encode.

 

[soresu]

While better speed at equal quality obviously makes huge sense for netflix, still the streaming quality is usually pretty bad. I mean youtube 1080p can easily look much worse than a good 720p encode.

Depends on the channel, I have my suspicions that some are given priority there by paying for the privilege - the Playstation channel often has flawless encodes.