Aren't yields also a function of voltage as well as clockspeed? It's not just a question of how many chips simply work but also how many work at the set clockspeed and thermal parameters?
That's what I meant from mentioning the following:
AlStrong said:it's not just whether the chip works or not, but also getting chips at the desired clockspeeds (or rather, chips that behave within operational power and thermal designs)
Power consumption can behave differently at the same voltage depending on, for example, gate design, the doping or just a synergy of random defects that make the chip crappier than others.
You mean on the same node? Potentially, but they'd probably want to play it safe and easy with just a single production configuration until a major revision. They'd also be able to skip testing chips as though they were binning them when that's not an actual consideration. Probability of chip failure is increased with higher temperatures, so they might as well just stick to the one type of heatsink that covers the worst case (IIRC, MS's TRC for randomly sampled units was 36 hours operating under load before failure).So if worse comes to the worse they could go with say a vapour chamber cooler initially and then spec it down to a regular heatsink/fan combo when yields improve and required voltages drop?
Might be too supply constrained. IIRC, Global Foundries isn't really well positioned for mass production until 2012.One burning question I have is whether they could possibly get a console out in 2011 as those rumours suggest or early 2012 even on 28nm since we now have two contract fab companies slated to offer 28nm production this year? Thats the real question since it would have an extreme bearing on what they could pack in at their desired thermal limit. Although in saying that 32nm SOI is also a possible candidate although I don't know how much capacity GF can spare to anyone who isn't called AMD!
Plus we still haven't even seen desktop graphics cards on 28nm from AMD/nVidia, and these two have lower scale production i.e. they've never required 1-2 million working chips at launch.
One would love smaller than 28nm, but that should be quite mature by 2013.Anyway, which node do you think?