David Manheim, ALTER founder, argues physical cooling limits will halt exponential AI hardware scaling by the early 2030s
Anders Sandberg suggested hierarchical nanochannel cooling as a potential workaround.
@davidmanheim The deep limit is when electrons go unbound or chemical bonds shake apart; this is actually far above a fission nuclear reactor temperature. But I don't think doped silicon handles high temperature well, so there is definitely a bound.
David Manheim@davidmanheimIt's usually smart to extrapolate exponentials for AI instead of positing we'll hit the top of the S-curve. But I'd reasonably expect an exception when we start discussing when theoretically known cooling methods can't manage the power density - in this case, early 2030s.
@davidmanheim @snewmanpv OK, current (early afternoon) estimate is 1 MW - 10 MW per cubic meter, with upper edge involving exotic coolants. This makes some optimistic assumptions about rack density that might make it a factor 5 too high? Anyway, lower than nuclear core power density.
Anders Sandberg@anderssandberg@davidmanheim @snewmanpv On earth conduction and convection tends to win, but I wonder if one could do something like the super-Planckian radiators to boost transfer...