The Intel SVT-VP9 Performance Boost Across 10 Intel/AMD Systems

[malakudi]

Michael, can you run 7980XE with AVX512 disabled to see the gains AVX512 offers over AVX2 on the same CPU?

[milkylainen]

H264 is already covered by HW pretty much anywhere (or achieves "good enough" speed in existing SW, it's hard to beat x264).

I don't agree. It's like saying that gzip speed and compression suits everyone.
x264 is pretty slow, even in crap settings.
But x264 is actually far better of than hardware encoders when it comes to quality.
I would use x264 over hardware encoders if I was encoding data for h264 storage.
But for realtime live streams use on a low power CPU, x264 is just to slow.
Faster options with quality tradeoffs in software is a nice thing to have.
Besides, h265 is patent encumbered in ways h264 is not.

I still maintain that a h264 addon to the SVT suite would be a great thing.

[Michael]

So basically the test is worthless is what you just said. The first pass is the analysis pass, the second pass is the one that does the actual encoding and as anyone that has ever done any amount of encoding knows, it's very common for the first pass of many encoders to be 2-4 times faster than the second pass.

Further, since many times you may only do a 1 pass encoding, such as using CRF or QP modes or CBR encoding, the analysis pass is "folded" in with the encoding pass.

Your tests are very misleading IMHO, all this time I was under the impression that you were testing a single pass encoding scenario, knowing that Intel sped up the first pass makes these results a lot less impressive.

The SVT-VP9 test profile reports an overall FPS, that is what's used for performance, not any first pass. AFAIK, libvpx is the only one not exposing an overall metric and thus just using the first pass.

[Spooktra]

The test profile measures currently first pass, the second pass is slower.

So basically the test is worthless is what you just said. The first pass is the analysis pass, the second pass is the one that does the actual encoding and as anyone that has ever done any amount of encoding knows, it's very common for the first pass of many encoders to be 2-4 times faster than the second pass.

Further, since many times you may only do a 1 pass encoding, such as using CRF or QP modes or CBR encoding, the analysis pass is "folded" in with the encoding pass.

Your tests are very misleading IMHO, all this time I was under the impression that you were testing a single pass encoding scenario, knowing that Intel sped up the first pass makes these results a lot less impressive.

[tildearrow]

The test profile measures currently first pass, the second pass is slower.

Oh! That explains.

[Michael]

I didn't think libvpx-vp9 was that fast...?

Maybe try -O3 on SVT?

The test profile measures currently first pass, the second pass is slower.

[tildearrow]

I didn't think libvpx-vp9 was that fast...?

Maybe try -O3 on SVT?

[flashmozzg]

SVT is one of Intels ongoing efforts to make "Software instead of Hardware" solutions to problems.
If they were serious about it they would add H264 already.
It would make SVT realtime encoding a possibility on a bucketload of CPUs today without eating the CPU to bits as in the h265 case.

So pretty please Santa. With sugar on top. Will you bring H264 to the SVT-suite?

H264 is already covered by HW pretty much anywhere (or achieves "good enough" speed in existing SW, it's hard to beat x264).

[Nille_kungen]

I was hoping to see some AMD APU in the mix, but still i did appreciate the testing in this article.

[Jumbotron]

My Bristol Ridge based laptop and desktop are happier now. Just as a reminder....AVX 2, otherwise known as HNI or Haswell New Instructions were introduced into said Intel's Haswell line of CPU's and continue to this day into Ice Lake.

However, AVX 2 only began to show up in AMD processors in 2015 in the Excavator CPU's which were the last of the Bulldozer line. So that would be.......

Carrizo APU
Bristol Ridge APU
Stoney Ridge APU
Toronto APU

And four desktop Carrizo based CPU's with their integrated GPU's fused off.

Athlon X4 845 ( based on Carrizo )
Athlon X4 940 ( based on Bristol Ridge )
Athlon x4 950 ( based on Bristol Ridge )
Athlon X4 970 ( based on Bristol Ridge )

Of course...AVX 2 has been in ALL the Zen processors from the start.