A petaflop here, a petaflop there, and pretty soon you are talking about real supercomputing

NSF formally announced their awards which other had hinted at over the past few weeks. These machines will be “500x faster than todays supercomputers”. How this will occur in 5 years, is well, not know. Moore’s law (if it holds) gives us an order of magnitude in 5.5 years or so. So thats 50x faster than Moore’s law following units.

Your options here are a) make more of them, or b) make em faster. Or all of the above. More of them can be done, though we are within an order of magnitude of the thermal deBroglie \ wavelength of electrons at 300K. Quantum effects exist today, they will likely dominate as we get down to those size features. Would be fun to calculate what happens to the conductivity of the “wires”. Shape effects get very important down there. Could create all sorts of interesting resonance levels with accidents of manufacturing. Make em faster usually involves die shrinks. Material switches are hard, and very expensive. Assume Silicon for the forseeable future. It is cheap, plentiful, and doesn’t have a high LD50. Not like GaAs … the material of the future (and always will be). I still think we need to stop using fermions to compute with (electrons, holes, ..). Bosons are needed. Naturally (massively) parallel by design. Building an optical ALU still may be an issue. Well, more power (and cooling) to these folks. They are going to need it. So the important question is, which language will it be programmed in, and how efficient can we get it?