Computational Requirements for Human-Level Intelligence



   No one really knows at this time just what computational capabilities will be required for full human-level intelligence. Ray Kurzweil has estimated it at 20 quadrillion calculations per second. I've speculated that he may base his estimate upon the fact that there are of the order of a quadrillion synapses in the brain. The brain activates them about 20 times per second. One quadrillion times twenty is twenty quadrillion. However, the activation of a synapse is probably not equivalent to a calculation. I favor Dr. Hans Moravec's estimation.. Dr. Moravec notes that, based upon his extensive personal experience, it would require about one billion calculations per second to emulate the four retinal layers that pre-process optical signals before sending them to the brain. These four retinal layers occupy a 2 cm. diameter. X 0.01 cm. volume in the retina, or about 0.03 cms3. Comparing this to the brain, with its volume of 1,350  , yields a ratio of about 40,000. On this basis, Dr. Moravec estimates that the brain must be able to process of the order of 100 trillion calculations per second, or about 100 times as fast as the 2005 "Cell" microprocessor, or about about 5-to-10 times what I'm guessing for the speed of the fastest 2010 "Cell" chip. (Dr. Moravec uses 0.02 cms. for the volume of the retinal processing layers, and 1,500 cms.for the volume of the brain, and comes up with 70 trillion operations per second, which he rounds up to 100 trillion operations per second, but factors of 2 are a joke when you're looking at a factor of a 100,000:1.) In addition, I'm speculating that a mobile robot won't have to be as capable as a human being. It won't have to perform at the near-Olympic level of a well-trained athlete. Its vision, hearing, smell, taste, and touch may not have to be as capable as they are in a 21-year-od human. But the gist of the matter is that, if the "Cell" comes to fruition, we might be approaching hardware capabilities that can support high-level robotics sooner than we'd expected.
    Another factor that enters into these deliberations is that the cost of RAM isn't expected to drop faster than its traditional Moore's Law curve would indicate, and it also plays a crucial role in AI. Dr. Moravec is currently using ~500 megabytes of RAM
 
 
 

Computer power chart