The Robots Are Getting Closer - 7


Future Index

Hardware Cost and Capability As a Function of Time
    Last month, PC Progress would sell you gigabytes of 133 MHz SDRAM for $42, making it $84 a gigabyte, plus tax. gigabyte of faster 333 MHz SDRAM cost $124, or about three times as much...  about $250 a gigabyte. However, for some artificial intelligence functions, fast RAM may not be all that important.
    Disk storage ran as low as $139 for 120 gigabytes.

Table 1 - The Computing Resources That a University, and an Individual Might Afford


 University, $20,000/Year Budget

An Individual, $2,000/year Budget


TB of RAM 

Disk, TB Processors RAM, TB 
Disk, TB Processors
2002 0.06/0.2 14.4 Sixty 2.4-GHz P4s 0.006/0.02 0.96 Six 2.4-GHz Athlons
2005 0.24/0.8 60 10-20 Cells 0.02/0.075 4 1-2 Cells
2010 1/3 500 10-20@10-Tflps. 0.1/0.3 30 1-2@10-Tflps.
2013 3/10 1,000 300-600 Tflps. 0.3/1 150  1-2@30-Tflps. 

    As of 1/9/2003, a 2.4 GHz P4 with 1 gigabyte of 1,066 MHz RAMBus and two 0.08 TB hard drives sells for $1,078. (There may be ways to mount several motherboards in one large cabinet, but buying self-contained computers would permit their re-sale every year or two.) I'm assuming that these computers can be sold for one-third of their cost every two years.
    An alternative approach would be to buy two Apple 1.25 GHz dual-processor G4's for $3,294 each. (I believe that Dr. Moravec is using Apple computers for his research.) These are advertised to deliver 18.3 peak  gigaflops each, for a total peak processing speed of 36.6 gigaflops. They each come with 1 gigabyte of RAM and a 120 GB disk.
    It's worth mentioning that a reasonable projection for the successor to the G4 in
2005 would be 50-to-60 gigaflops
2010 would be 275-to-325 gigflops
2013 would 675-gigaflops-to-1 Teraflops. 
    This table assumes that both the university and the individual can sell their computers after two years for one-third of what they paid for them. 
    There may be better ways to go about this, but this might set a lower bound upon what should be available over the next 10 years.
    By 2013, an amateur may be able to afford the computing resources for human-level AI, with 30 to 60 teraflops of processing power, 0.3 terabytes of SDRAM, and 225 terabytes of disk in six computers.
    If Intel delivers a teraflops processor by 2010, they ought to be delivering 3 teraflops chips by 2013, permitting the dedicated amateur to wield 18 teraflops of general purpose processing power. Not bad for a fallback position, in case the Cell turns out to be a paper tiger! (I've predicted 600 gigaflops for the mainstream Intel processor for 2013, but Intel has said that it will deliver a teraflops chip by 2010, so we'll see.)
    Extrapolating to 2015, we would expect 5-teraflops Intel processors. 
    Extrapolating further, we might expect 15-teraflops-per-chip on tap for 2018 (for a total of 90 teraflops). Actually, Intel might mount more than one processor on each chip, boosting the output more than I'm predicting. 
    Human-level-AI software development will really have to hump to keep up with this hardware schedule!

    The above table is reproduced in greater detail below.

    A university robotics research program might have been able to afford 

[* -  I question this. Ten gigaflops out of a 2.5 GHz Pentium 4 sounds too high to me, but that's what one of the articles quoted. An Apple 1.25 GHz dual-processor G4 cranks out 18.3 gigaflops.]

    A dedicated individual might have been able to afford 

    A university robotics research program might be able to afford:

    A dedicated individual might be able to afford

    A university robotics research program might be able to afford:

    A dedicated individual might be able to afford:

    A university robotics research program might be able to afford:

    A dedicated individual might be able to afford:

    Another important development will be the availability of extremely-wide-bandwidth (480 Mbits/second), wireless data links. Apple computer has just announced that its new laptops will employ the 54-megabit-per-second 802.g wireless standard, and a 480-Mbit/second standard is in the works. This will mean that desktop computers can receive and process raw visual inputs from a mobile, visually-navigating robot. The robot won't necessarily require a lightweight, low-powered laptop for onboard operation.

    If the Cell processor is inadequate to perform the required calculations, a Pentium 4 or its equivalent should deliver:
2002:    10 gigaflops
2005:    40 gigaflops
2010:  200 gigaflops
2013:  600 gigaflops

    Although this is far below 100 teraflops, several conventional processors might be harnessed even by a private individual to bring computations into the teraflops realm by 2013, and universities could probably muster the money to reach more than 10 teraflops by that date.

Back                                  1   2   3   4   5   6   7   8   9   10   11   12                                   Next