Robots Are Getting Closer - 12
The Long-Term Roles of Robots
We've Been Living in the Age of Automation and Robotics for 50 Years
We have been living in an "Age of Robotics" for 50 years, with multifarious changes already arising from it. Part of today's relatively low cost of manufactured goods may be a result of manufacturing automation.
How Much Information Can the Human Brain Hold?
The human brain, with a volume of the order of 1,300 cc's., has about 100 billion neurons, and 5,000 to 10,000 times as many synapses. If we suppose that a synapse can store something like a byte of information (256 levels), then the human brain would be capable of storing of the order of one quadrillion bytes (1,000,000 gigabytes) of information. (I think that this might be a little on the high side. For one thing, synapses would be an unreliable form of storage, and more than one synapse might be required for dependable storage of a byte of data.)
What Would It Cost to Duplicate This Much Storage Today?
Note that this method of estimating the computer-capacity equivalents of animal brains is independent of the calculation-speed approach, although it turns out to track the calculation-speed approach fairly well.
At this time, a terabyte (1,000 gigabytes) of disk storage would cost $1,250, with prices dropping to, perhaps, $625 by year's end. A quadrillion bytes of RAM would cost about $1,250,000... doable, but not very practical for mowing the yard. If 100,000 gigabytes will do the job, it would cost $125,000, dropping to, perhaps, $62,500 by year's end... still a little high for yardwork.
A gigabyte of RAM costs $84 to about $320, depending upon its speed, with ten gigabytes costing $840 to $3,200..Of course, 10 terabytes would set us back $840,000 to $3,200,000. If we could get by with one terabyte, it would only run $84,000 to $320,000.
How Much Memory Could We Afford Today?
Today, we might economically afford to install, perhaps, a terabyte of disk storage and eight gigabytes of RAM in an industrial robot. This would be about 1/1,000th the presumed storage capacity of the human brain, and would correspond to an animal brain containing 100 million neurons. Such a brain would have about 1/1,000th of the volume of the human brain, and would correspond to 1/1,000th of the brain capacity of the human brain, or a cranial endowment of about 1.3 cubic centimeters.
Such a brain might belong to a small mammal.
Of course, these are seat-of-the-pants estimates and could easily be off by a factor of ten in either direction.. The true computing requirements may not be known any time soon.
Comparing the Storage-Cost Approach with Dr. Moravec's Calculation-Speed Approach
Looking at the calculation-speed approach, 1/1,000th of Dr. Moravec's 100 trillion instructions per second that he suggests is representative of the speed of the human brain leads us to a computing speed of 100 gigops or 100,000 MIPS. A 1.25 GHz dual-core P4 can crank out about 14 gigaflops, requiring seven chips to achieve 100 gigaflops... again, within current price range. And that's about what an industrial installation might be able to afford today.
Ten years ago, in early 1993, RAM cost about $30 a megabyte ($30,000 a gigabyte), or about 100 times today's price, and disk storage cost about $0.50 a megabyte ($500,000 a terabyte), or about 400 times today's price). Consequently, in 1991 we could have afforded to install in an industrial robot about 1/100th of the storage capacity that we could afford today, corresponding to a brain volume of about 13 cubic millimeters--the brain volume of, perhaps, a large insect, or a small lizard.
Early 90's Computers Had the Storage Capacities of the Brains of Large Insects
Computing speeds would have been far below this in January, 1993, running at, perhaps, 20 MIPS. Five such chips would have only yielded 100 MIPS, or about 1/1,000 the output of seven current P4 chips. And we're comparing MIPS with gigaflops here. These 1993 chips would compare even less favorably if we considered their floating point speeds.
This agrees well with Dr. Moravec's "insectile" intelligence for 1980's and 1990's robots.
It Will be 2015 to 2022 Before Human-Level Industrial Robots Are Affordable
Projecting ahead, to the extent to which these assumptions are valid, it will take about 20 years of Moore's-Law improvements before memory for an industrially-affordable robot brain can rival the hypothetical 1,000,000 gigabyte storage capacity of the human brain, together with . (This will require circuit features of 11, or possibly, 16 nanometers.) This chain of reasoning would lead us to high-priced robots in 2022.
It might take another eight years (2030) to bring storage costs down to consumer levels.
If we could get by with 1/10th of these capacities, we might get to human-level industrial robots by 2015.
9 10 11
12 13 14