Hacker News Reader

Comment by simias | Hacker News Reader

Having worked on the particular issue of streaming real-time video over wireless, I can say that the main issue is not link latency.

The main lag comes from encoding/decoding. If you do it naively you encode frame-per-frame (encoding slices is more difficult), and the encoder does not only outputs iframes: you get partial frames that depend on both previous and future frames. Also the decoder does not always output frames in order. So you have to expect something around ~10 frames of latency, maybe less if you optimize everything well enough. That still means easily more than 100ms of lag.

The network is really not the main issue here.

View on HN

I would think that the people at Valve would be able to find a solution if anyone could... or are you claiming that this is a hard nut to crack and SteamOS's streaming solution won't end up being that great for high-resolution TVs?

Oh I'm sure if they put their minds to it they can make some improvements and clever optimizations. The thing it becomes exponentially more difficult the lowest the latency you want to achieve, obviously. <300ms? easy. <100ms? manageable. <50ms? hey, very good! <10ms? uh, I want to see it with my own eyes.

Also, you have to remember that steam won't control the encoding end of the pipeline (and if the "steambox" is third party hardware with steamOS installed, no control at all on the hardware). Which means that in the end the observed latency will depend a lot on the hardware and drivers of the desktop PC and there isn't much Valve can do about that.

So in the end I'm sure it'll be more than fine to play Civilization or Torchlight, MMOs and most RPGs but maybe not Counter Strike or Quake III.

Why are people acting like this is impossible even after it has been done? Remember onlive? Notice how the latency was entirely the same as your network latency to their servers, and there was no problem with encoding adding any (noticable) additional latency?

What about scrapping the conversion to streaming video entirely and replacing it with a networked graphics protocol that allows one PC to draw on another's graphics natively?

It seems to me, intuitively, that the only ways to do that are:

1. Essentially equivalent to streaming, or

2. Essentially equivalent to normal CPU->GPU communication but over the network (and, thus, needing the bandwidth that local CPU->GPU communication has if you want to avoid slowing things down -- GigE wouldn't seem to be enough, much less WiFi, even if you consider only bandwidth and not latency.)

Well that would simplify things greatly of course, but then it means bit hit on the bandwidth.

I mean, a 720p60Hz stream in 4:2:0 (12 bits per pixel) still amounts to 663Mbits/s. You won't get that out of a gigabit link realistically (at least not over IP). Of course you could use a lightweight compression algorithm, but you'll have to divide this bandwidth by at least 5 to make it manageable for the average home network I'd say (I have absolutely nothing to back that last number, but 100Mbits/s doesn't look too scary...). And that's only for 720p remember.

I think if you plan to stream HD video over the network you have to encode and be clever about it.

>You won't get that out of a gigabit link realistically (at least not over IP)

I just got more than that using scp with no compression. Yes, you absolutely can get that realistically from gigE.

Reply on news.ycombinator.com