Re: Blue Gene

Robert J. Bradbury (
Mon, 6 Dec 1999 10:21:57 -0800 (PST)

On Mon, 6 Dec 1999, Dan Hook wrote:

> According to the Reuters article on the subject, this computer will run at
> 10^15 ops per second, which is only two orders of magnitude below the
> 10^17 ops per second that various people project as the computational
> equivalent of the brain.

Actually, if you take Moravec's numbers a brain equivalent computer is only about 10^14 OPS. If you simply look at the neocortex, which is to a large degree makes humans more intelligent than say animals, you have 20 billion neurons (20*10^9) and since they can only fire at a rate of 10-100x/sec, the neocortex OPS rate is more like 10^11-10^12 OPS (which starts to explain how you can do speech recognition on much less than giga-op computers).

> It seems that we are heading towards a situation where the hardware is
> willing, but the software is weak.
No, you have neural net software. Its the architecture that is holding you back currently. The thing the brain is good at is having *many* processors and *lots* of communications. But it looks like the computer architects are finally getting it right (lots and lots of really simple processors that communicate pretty well).

You should look at this development as very good for the prospects for AI, it makes relatively little difference for hard nanotech and is pretty much irrelevant for programmable (wet) self-replicators since we will have those before this architecture makes it out of the lab.