That said the 20- series is really freaking expensive. When I got my C1060's, they were ~$1500/per retail; I think the C2070's are nearly twice that.
That might be hard when going for reliability. It's a bit like overclocking your CPU, you can do it, but it's outside of the warranty. In this case the overclockers are the third party board manufacturers that are trying to get an advantage over their competitors that they can paste on the box so that in the store when comparing two boxes the customer will pick theirs.
> (Also, the value of ECC is diminished if a Tesla is over 3x the price of a GeForce, because then you can run three times and vote.)
That's not really true. The majority of the situations where the Tesla boards are used you'll find multiple boards in one box already. The limiting factor here is how many boards you can pack in to a system, more = better. So if you'd use the 'voting' principle you'd be wasting tons of power, rackspace and host computers in order to offset the price of the ECC, not just the two extra cards.
I don't think so; most GeForce cards appear to use stock clocks and Nvidia reference PCBs and coolers.
I expect we'll see the exact same thing happen as what happened with the 200 series, the first cards will be looked at as the crummy old ones as soon as the next batch hits the streets. You can pretty much expect the following:
a 1.4 GHz part
maybe a 1.8 or 2.0 GHz part
a dual chip card featuring 896 cores at the same power consumption level as the current 480GTX