Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!usc!zaphod.mps.ohio-state.edu!wuarchive!emory!gatech!rutgers!aramis.rutgers.edu!athos.rutgers.edu!nanotech From: landman@eng.sun.com (Howard A. Landman) Newsgroups: sci.nanotech Subject: Re: Down and Out in Nanoland Message-ID: Date: 21 Jan 91 18:36:27 GMT Sender: nanotech@athos.rutgers.edu Organization: Sun Microsystems, Mt. View, Ca. Lines: 37 Approved: nanotech@aramis.rutgers.edu >[It'll take at least a MIPS to simulate a synapse with any claim > to fidelity, so we need 10 trillion MIPS (that's 10 million tera-ops > or 10^19 instructions per second) to run the simulation.] Except that this makes the rather stupid assumption that we'll prefer to simulate a brain rather than build one. Simulating a million-transistor VLSI chip takes perhaps 10^7 times as long as just running it directly. This would indicate that only 10^12 ips (a million MIPS) might be enough if they're the right kind (custom hardware). Carver Mead's analog neural chips (the artificial "cochlea" and "retina") use around 1 to 3 transistors per synapse-equivalent and operate many orders of magnitude faster than biological ones. > At an absolute minimum we must consider 10 billion neurons, with > an average 1000 synapses each, which can fire at rates of a kilohertz. So either we build a 10 trillion transistor brain which will operate at least 1000 times faster than a human one; or we figure out ways to use fewer neurons switching faster to emulate more neurons switching slower ("virtual neurons") which gets us down to perhaps a billion transistors. Even in today's technology, it's conceivable to pack that much circuitry in the volume of a tunafish can and easy to fit it into a breadbox. With density doubling every 3 to 4 years, we should be able to build a brain equal in power and size to a human brain within a decade or two. Assuming that we knew how to design it, of course. -- Howard A. Landman landman@eng.sun.com -or- sun!landman [This is all reasonable but the argument was about whether it was going to be easier to upload or AI. The assumptions were to separate the sides of the argument, not prescriptions about the best way to achieve human-level intelligence in a machine. --JoSH]