3 ms·

> To me, the big news here is that ~30% of devices now support AV1 hardware decoding Where did it say that? > AV1 powers approximately 30% of all Netflix view

by 0manrho 11mo ago

> To me, the big news here is that ~30% of devices now support AV1 hardware decoding Where did it say that? > AV1 powers approximately 30% of all Netflix viewing Is admittedly a bit non-specific, it could be interpreted as 30% of users or 30% of hours-of-video-streamed, which are very different metrics. If 5% of your users are using AV1, but that 5% watches far above the average, you can have a minority userbase with an outsized representation in hours viewed. I'm not saying that's the case, just giving an example of how it doesn't necessarily translate to 30% of devices using Netflix supporting AV1. Also, the blog post identifies that there is an effective/efficient software decoder, which allows people without hardware acceleration to still view AV1 media in some cases (the case they defined was Android based phones). So that kinda complicates what "X% of devices support AV1 playback," as it doesn't necessarily mean they have hardware decoding.

sophiebits 11mo ago
“30% of viewing” I think clearly means either time played or items played. I’ve never worked with a data team that would possibly write that and mean users. If it was a stat about users they’d say “of users”, “of members”, “of active watchers”, or similar. If they wanted to be ambiguous they’d say “has reached 30% adoption” or something.
- 0manrho 11mo ago
  Agreed, but this is the internet, the ultimate domain of pedantry, and they didn't say it explicitly, so I'm not going to put words in their mouth just to have a circular discussion about why I'm claiming they said something they didn't technically say, which is why I asked "Where did it say that" at the very top. Also, either way, my point was and still stands: it doesn't say 30% of devices have hardware encoding.
- csdreamer7 10mo ago
  I am not in data science so I can not validate your comment, but 30% of viewing I would assume mean users or unique/discreet viewing sessions and not watched minutes. I would appreciate it if Netflix would clarify.
endorphine 11mo ago
In either case, it is still big news.
cogman10 11mo ago
That was one of the best decisions of AOMedia. AV1 was specifically designed to be friendly for a hardware decoder and that decision makes it friendly to software decoding. This happened because AOMedia got hardware manufacturers on the board pretty early on and took their feedback seriously. VP8/9 took a long time to get decent hardware decoding and part of the reason for that was because the stream was more complex than the AV1 stream.
- Neywiny 11mo ago
  Hmmm disagree on your chain there. Plenty of easy hardware algorithms are hard for software. For example, in hardware (including FPGAs), bit movement/shuffling is borderline trivial if it's constant, while in software you have to shift and mask and or over and over. In hardware you literally just switch which wire is connected to what on the next stage. Same for weird bit widths. Hardware doesn't care (too much) if you're operating on 9 bit quantities or 33 or 65. Software isn't that granular and often you'll double your storage and waste a bunch. I think they certainly go hand in hand in that algorithms relatively easier for software vs previously are easier for hardware vs previously and vice versa, but they are good at different things.
  - cogman10 10mo ago
    I'm not claiming that software will be more efficient. I'm claiming that things that make it easy to go fast in hardware make it easy to go fast in software. Bit masking/shifting is certainly more expensive in software, but it's also about the cheapest software operation. In most cases it's a single cycle transform. In the best cases, it's something that can be done with some type of SIMD instruction. And in even better cases, it's a repeated operation which can be distributed across the array of GPU vector processors. What kills both hardware and software performance is data dependency and conditional logic. That's the sort of thing that was limited in the AV1 stream.
    - IshKebab 10mo ago
      > Bit masking/shifting is certainly more expensive in software, but it's also about the cheapest software operation. In most cases it's a single cycle transform. He's not talking about simple bit shifts. Imagine if you had to swap every other bit of a value. In hardware that's completely free; just change which wires you connect to. In software it takes several instructions. The 65 bit example is good too. In hardware it makes basically no difference to go from 64 bits to 65 bits. In software it is significantly more complete - it can more than double computation time. I think where software has the advantage is sheer complexity. It's harder to design and verify complex algorithms in hardware than it is in software, so you need to keep things fairly simple. The design of even state-of-the-art CPUs is surprisingly simple; a cycle accurate model might only be a few tens of thousands of lines of code.

by 0manrho 11mo ago

sophiebits 11mo ago

“30% of viewing” I think clearly means either time played or items played. I’ve never worked with a data team that would possibly write that and mean users. If it was a stat about users they’d say “of users”, “of members”, “of active watchers”, or similar. If they wanted to be ambiguous they’d say “has reached 30% adoption” or something.

0manrho 11mo ago
Agreed, but this is the internet, the ultimate domain of pedantry, and they didn't say it explicitly, so I'm not going to put words in their mouth just to have a circular discussion about why I'm claiming they said something they didn't technically say, which is why I asked "Where did it say that" at the very top. Also, either way, my point was and still stands: it doesn't say 30% of devices have hardware encoding.
csdreamer7 10mo ago
I am not in data science so I can not validate your comment, but 30% of viewing I would assume mean users or unique/discreet viewing sessions and not watched minutes. I would appreciate it if Netflix would clarify.

endorphine 11mo ago

In either case, it is still big news.

cogman10 11mo ago

That was one of the best decisions of AOMedia. AV1 was specifically designed to be friendly for a hardware decoder and that decision makes it friendly to software decoding. This happened because AOMedia got hardware manufacturers on the board pretty early on and took their feedback seriously. VP8/9 took a long time to get decent hardware decoding and part of the reason for that was because the stream was more complex than the AV1 stream.

Neywiny 11mo ago
Hmmm disagree on your chain there. Plenty of easy hardware algorithms are hard for software. For example, in hardware (including FPGAs), bit movement/shuffling is borderline trivial if it's constant, while in software you have to shift and mask and or over and over. In hardware you literally just switch which wire is connected to what on the next stage. Same for weird bit widths. Hardware doesn't care (too much) if you're operating on 9 bit quantities or 33 or 65. Software isn't that granular and often you'll double your storage and waste a bunch. I think they certainly go hand in hand in that algorithms relatively easier for software vs previously are easier for hardware vs previously and vice versa, but they are good at different things.
- cogman10 10mo ago
  I'm not claiming that software will be more efficient. I'm claiming that things that make it easy to go fast in hardware make it easy to go fast in software. Bit masking/shifting is certainly more expensive in software, but it's also about the cheapest software operation. In most cases it's a single cycle transform. In the best cases, it's something that can be done with some type of SIMD instruction. And in even better cases, it's a repeated operation which can be distributed across the array of GPU vector processors. What kills both hardware and software performance is data dependency and conditional logic. That's the sort of thing that was limited in the AV1 stream.
  - IshKebab 10mo ago
    > Bit masking/shifting is certainly more expensive in software, but it's also about the cheapest software operation. In most cases it's a single cycle transform. He's not talking about simple bit shifts. Imagine if you had to swap every other bit of a value. In hardware that's completely free; just change which wires you connect to. In software it takes several instructions. The 65 bit example is good too. In hardware it makes basically no difference to go from 64 bits to 65 bits. In software it is significantly more complete - it can more than double computation time. I think where software has the advantage is sheer complexity. It's harder to design and verify complex algorithms in hardware than it is in software, so you need to keep things fairly simple. The design of even state-of-the-art CPUs is surprisingly simple; a cycle accurate model might only be a few tens of thousands of lines of code.