[ExI] The second step towards immortality
anders at aleph.se
Tue Jan 7 12:11:52 UTC 2014
On 2014-01-07 07:42, Andrew Mckee wrote:
> On Tue, 07 Jan 2014 05:33:43 +1300, Anders Sandberg <anders at aleph.se>
>> The transition from no human emulations and a world with emulation
>> may also be gradual if it is computing-limited: at first you need
>> major supercomputers to run an emulation in realtime, later Moores
>> law allow the big ones to run faster or multiple cheaper realtime
>> emulations, followed by ever faster and more numerous copies. In the
>> scanning- or neuroscience-limited transition the need for planning
>> ahead is larger.
> Ummm, but isn't Moore's law only two or 3 process shrinks away from
> being stone cold dead?
Haven't people been saying that since day one? Actually, *Moore* did it
in his original paper if I remember right. I have stopped listening to
detailed worries like this (sorry, Eugene) and instead have turned to
the data: when I fit logistic curves (implicitly assuming a stop) to
flops/$/s the lower end of the 99% confidence interval is five orders of
magnitude more powerful than today.
A world with merely 100,000 times better computers than today is likely
a no-upload world: at least detailed biophysics is not going to be
feasible outside giant installations. But I only give it a 1% chance of
happening outside catastrophe scenarios.
> So how many super computers can this future energy grid support?
Depends on the other law in town, Koomey's law
Up until recently energy was not much of a design criterion, so I
suspect we might even see an acceleration as we start caring more about
energy than size. The limits are as always tricky to judge:
> The current IBM roadrunner does 376 million calculations per watts. If
> we take my mid-range estimates of computing needs, 10^22 to 10^25
> FLOPS, then a single emulation would need 10^13 to 10^16 watts. The
> total insolation of Earth is about 10^17 watts, so this won't do -
> there would be space for just a few minds on the entire planet. But
> current research on zettaflops computing suggest we can do much
> better. A DARPA exascale study suggests we can do 10^12 flops per
> watt, which means "just" a dozen Hoover dams per mind.Quantum dot
> cellular automata could give 10^19 flops per watt
> putting the energy needs at 200-2000 watts.
And reversible computations are way better, of course.
Dr Anders Sandberg
Future of Humanity Institute
Oxford Martin School
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the extropy-chat