Computer Technology Forecast, 1997 – 2012 (Based Upon the Semiconductor Industry Association’s 15-Year Technology Roadmap)

Robert N. Seitz

September 30, 1997

Table 1: What You Might Expect on Your Desktop for $1,000-$1,200, 1997 through 2012:
 
Mo./
Year		 CPU
MHz/Type		 Speed
MIPS		 RAM
MB		 HD
GB		 Accel, 
Gigs		 CD, 
GB		 Comm, 
MBaud		 Video-camera,
Removable Disk?
4/97 133/Pent 160 32 3 1-2 0.6 0.033  
4/98 266/Pentium II 320 64 5 2-3 0.6 0.056   
4/99 400/Deschutes 800 128 8 5 4.3 0.112  
4/00 600/Katmai 1,200 128 12 8 7.6 0.256 640X480 Vidcam, Zip 100
4/01 900/Willamette 2,700 256 20 18 15.0 0.256 800X600 Vidcam,RW CD
4/02 1,200/Merced II 5,000 512 35 32 15.0 0.400 1,024X758Vidcam,RW CD
4/03 1,600/Merced II 7,000 1,024 60 45 15.0 0.4 1,280 X 960 Vidcam, RW
4/04 2,200/P8 10,000 1,024 100 65 60.0 1.0 1,920X1,080 Vidcam, RW
4/05 3,000/P8 12,500 2,048 200 80 180.0 1.5 System on a chip? DVD?
4/06 4,000/P8 II 16,000 4,096 400 100 180.0 1.5 System on a chip? DVD?
4/07 5,000/P9? 25,000 4,096 700 160 300.0 6.0 System on a chip? DVD?
4/08 6,500/P9 II?ium  40,000 8,192 1,000 250 500.0 6.0 System on a chip? DVD?
4/09 8,000/P9 II? 50,000 16,384 1,500 320 750.0 12.0 System on a chip? DVD?
4/10 9,500/P10? 80,000 16,384 2,000 450 1,000.0 12.0 System on a chip? DVD?
4/11 11,000/P10 100,000 32,768 2,500 650 1,000.0 12.0  
4/12 13,000/P10 120,000 32,768 3,000 800 1,500 12.0  

Notes for interpreting Table 1: (Column 2):Multiprocessing (use of multiple computers) is moving into the mainstream, with the future possibility of placing more than one microprocessor on a chip or of paying somewhat more for multiprocessor boards (which would sell more chips). This can provide an alternative method of achieving higher speeds for certain classes of partition-able, computationally demanding calculations. Also, the long-term role of Intel’s IA-64 (P7) architecture on the desktop is far from clear to me. Designations like Merced, P8, P9, and P10 are meant to be suggestive rather than prophetic. (Column 3): Computer speeds have become much more difficult to quantify with the introduction of super-scalar, super-pipelined, out-of-sequence processing, as well as MMx. Add to that changes in benchmarking from Megops to MIPS to SPECInt92s toSPECInt95s and it becomes hard to know where you are. Integer applications that are conducive to MMx may run several times faster than shown in Column 3. (Column 4) RAM (Random Access Memory) may move from EDO RAM in ’97 to SDRAM in ’98 and ’99 to Rambus in 2000. Access to memory has become the principal bottleneck to faster computing, and new techniques are appearing to facilitate faster and faster memory access. (Column 5): Hard drive capacities are complicated by the fact that:

  • (a) hard drive capacities have begun to double every year rather than every 2.25 years, and
  • 3.5" disk capacities are expected to top out at a theoretical limit of 500-to-1,000 Gigabytes.
  • I’m assuming that won’t happen. (Seagate has just announced a new laser-optical registration approach currently used for "flopticals" that is touted to economically increase hard drive capacities by 10-to-20-fold, permitting, by 2012, 20-terabyte drives at, perhaps, $40-$50 a terabyte.)
  • (Column 6) Multimedia accelerators may be absorbed into "system-on-a-chip" processors, first for low-end systems by 1998-99, and later, for high-end computers. However, MMx2 should permit digital signal processing applications to run 6 or 7 times as fast as the main computer.
  • Column 7) 160-GB CD drives are in the works for 2005, with 1,000-GB CD drives forecast for 2010.
  • (Column 8) Communications bandwidths are a bit of a wild card, depending as they do upon information-utility politics and upon upgrading of Internet servers, lines, and hubs. Modems are switching to software implementations. Sound cards will probably follow.
  • (Column 9) Video cameras are expected to drop as low as $30 by 12/98. Some sort of successor to the floppy disk is essential. Rewritable CDs may be the cheapest way to go.
    •
        "Systems on a chip" are forecast for the early years of the 21st century, in the interests of both speed and economy.

    Looking Ahead: What You Might Expect to Buy for $1,000 to $1,200
    April, 1997:

  • 133 MHz Pentium
  • 32 MB EDO RAM
  • 2.5 GB Hard Drive
  • Diamond Stealth 2000, 2 MB 3-D Graphics Accelerator
  • 16X CD Drive
  • 33.6 Fax Modem
  • 16-Bit Sound Card, Speakers, Microphone
    •  April, 1998: April, 1999: April, 2000: April, 2001:     I should emphasize that the above kind of detailed prediction is fraught with peril. For example, Computer Technology Overview
        We are in the midst of an ongoing computer technology revolution that dwarfs anything else in human experience. Today, you can buy a megabyte of RAM (Random Access Memory) for $3.00. Thirty years ago, in 1967, my employers at NASA paid $3,000,000 to buy that same megabyte of RAM for their Univac 1108 computers—a 1,000,000:1 price reduction! If this had happened in the automotive world, it would be as though you could buy a new Mercedes, costing $10,000 in 1967, for 1¢ today! Twenty years ago, in 1977, when Radio Shack and Commodore introduced the world’s first personal computers, Radio Shack had to charge $266 for 8 kilobytes of RAM. Today, 64 megabytes of RAM costs $192 (about a 10,000-to-1 ratio).
        It may well be that the computer sitting on my desktop is more powerful than all the world’s computers put together in 1967. Here again, there is a price/performance improvement of the order of 1,000,000:1. But if that doesn’t astonish you, try this. In 1946, ENIAC (Electronic Numerical Integrator and Calculator), the first electronic digital computer, was able to perform something like 3 decimal calculations a second. Fifty years later, in 1997, Intel delivered a computer that performs one trillion decimal calculations per second—300 billion times faster than ENIAC! A 10-trillion decimal calculation computer is forecast for 2000, rising to 30 trillion in 2001 and 100 trillion in 2005.
        The cheapest magnetic disks in 1967 stored one (1) megabyte of data and probably added $10,000 to the price of an IBM 1130 minicomputer. The current price is about $.04 a megabyte (Until recently, magnetic disk prices declined somewhat slower than RAM prices.) Appendix A contains plots and further discussions of these price trends.
        For the past 30 years, both speed and storage have improved by a factor of 10 every 5 years—Moore’s Law.
        There is reason to believe that this astounding rate of computer technology improvement will continue through at least the year 2012 and perhaps, through the year 2022. The Semiconductor Industry Association’s 15-year Technology Roadmap projects 13,000 MHz clock speeds (compared to 266 MHz today) and 256,000- megabit DRAM memory chips (compared to 256-megabit chips today) by the year 2012. Appendix B discusses the reliability of these forecasts. (I have been publishing computer technology forecasts since 1976. Appendix C contains these prior forecasts.)
        What will we do with all this speed and storage capacity?
        There is enormous room for improvement in computer capacity over the next fifteen years, just as there was 15 years ago. Computer games have begun to use multiple CD ROM disks. They could easily and very profitably utilize tens of gigabytes of DVD storage if it were available. By the year 2012, I predict that game developers will find it easy to fill up terabytes of DVD storage (giving us photo-realistic virtual worlds to explore!)
        The same enormous need for improvement exists with respect to RAM and hard disk capacities, and to computer speeds. Artificial intelligence, speech recognition, natural language understanding, computer vision, improved data compression, graphic/virtual reality/laser holographic displays, MPEG2 and MPEG4 encoding, and computer games and simulations can all profit mightily from orders-of-magnitude improvements in computer speeds. Right now, games like "Riven" have to jump from snapshot to snapshot rather than allowing virtual-reality type movement. Also, when "Riven" runs a "video clip", the resolution has to be reduced to blurry pixel blocks so that the computer can alter the image in real time. It could easily use a factor-of-ten improvement in computer speeds just to handle its video clips. A speed-improvement factor of 1,000 may not be enough to support photo-realistic virtual reality.
        In 1982, we were playing Pac-Man on our IBM and Commodore 64 computers. We might have wondered then how in the world we could use a 1,000-fold improvement in computer speeds and storage capacities. Now we know, and it has been wonderful. The next 15 years will be equally wonderful.
        Jesse Burst, the editor of the ZD Anchor-Desk has suggested that we may look back upon the year 2000 as the year in which computer age began for the consumer.

    Computer Applications Timeline
        It is even more dangerous for me to try to forecast future applications of computer technology then the technology forecasts themselves, since they depend both upon knowledge I don't possess, and upon non-technical factors. However, I think that, over the next few years, personal computers are going to impact the man in the street to an unprecedented degree. The next ten or fifteen years--and in fact, the next few years--should be very exciting. The timeline depicted below should be considered to be a listing of some possible near-term technology applications rather than a reliable schedule for these applications to appear.

                                       Voice mail                           Video-telephony?... Conversational dolls begin to appear
                                                                 Video mail.......Virtual reality becomes major entertainment medium?
    1st-gen........................... 2nd-gen.................................... Virtual shopping? ................................Voice-writers
    voice- ..............................voice- ......................................3-D displays? ...................................begin to replace
    writer............................... writer................... Net video-casting .......................................................typists
     
     
    1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

  •          Natural ....................................................................."Personal assistants" .............."Office Assistants"

    •          language ........................................................................begin answering ......................begin handling
             queries .............................................................................questions........................ routine secretarial
                                                                   Robotic lawn mowers, scrubbers, sweepers?............assignments
                                         Basic voice-operated language translation Fluent voice-operated language translation

    Figure 4 - Computer Applications Timeline

    These projections are discussed in greater detail in the following eight paragraphs.

    (1) Voice-Writers

  • The first consumer-oriented "voice writers" appeared in June of this year (1997) when Dragon Systems introduced "Naturally-Speaking", the first general-purpose, continuous-speech dictation system. It was priced at $695 ($169 by December), including a noise-canceling microphone. By September, IBM had its competing product, "ViaVoice", on the market for $99, including a noise-canceling microphone. ViaVoice requires a 166 MHz Pentium, preferably with MMx, 32 MB of RAM and 130 MB of disk space—too steep for most 1997 PC owners. The ability of these first "voice writers" to recognize speech is currently far inferior to human speech recognition, just as early chess-playing programs were far inferior to human chess players. (They make many mistakes.) However, as computer capabilities continue to increase and as a market develops for voice dictation, they may be expected to steadily improve.

    • Star-Trek-class speech recognition, when it is achieved, should permit the dictation of reports and letters faster and easier then can be done by keyboard. (Professional typists may become a vanishing breed.) Although many of us may find voice dictation a valuable tool from the outset, voice dictation may not begin to edge out manual typing for 5 to 10 more years.

    (2) Natural Language Understanding and Conversation with the Aid of Artificial Intelligence
        Natural language queries first appeared in 1997 with Microsoft Office. Natural-language understanding means that you can ask the computer questions in everyday English, such as "How do I double-space documents?" The computer (usually) correctly interprets this question and tells you how to do what you want to do. This capability is rumored to be a part of the Windows 98 operating system when it is introduced in 1998. One can probably expect to see it becoming widespread for "Help" functions in various software packages by next year (1998). The ability to frame sentences or to draw upon a large library of stored phrases may be expected as the next stage in this progression that will eventually lead to limited "understanding" in certain specialized fields. Computer-based telephone answering software that can process messages--e.g., giving a forwarding number to a select set of people--will probably come next, followed by software that will "understand" simple statements and queries, and will respond intelligently, permitting dialogs with computers. (Perhaps you’ll soon get an answer to the question, "Does my computer really love me?") However, before this can happen, computers will need to improve their voice recognition capabilities by becoming more accurate and speaker-independent. It will probably be the early years of the next century before this can happen—5 to 8 years from now.
        "Personal Assistants" may answer the telephone for us, and keep track of appointments, notifying us verbally.
        Much of our equipment may respond to the spoken word.

     (3) Voice-Operated Language Translation
  • Voice-operated language translation can probably happen right now, using high-speed laptop computers. Within months, a continuous-speech recognition system, such as IBM's "ViaVoice" or Dragon Systems' "Naturally Speaking" could be coupled with a good translation program and a foreign-language text-to-speech program to accept inputs from a microphone and deliver outputs through the laptop's speaker. You will speak into it in English and will see the sentence(s) printed on the screen. When you are satisfied that they are correct, you will OK them and the translator’s speaker will speak them in Arabic, Turkish or Japanese (a la Star Trek). Ultimately, this might take the form of a pocket-sized, clip-on device with a throat microphone and built-in speaker. It might store a very large number of common phrases such as "Do you have a Men's/Ladies' room?" or "I'm from Huntsville, Alabama", as well as translating extemporaneous inputs. You might even store your own custom library of translated phrases before you visited another country. Such hand-held devices may appear within one to three years.

    •     Microsoft and Apple are both working feverishly on speech input and output.     (4) Robotic Lawn Mowers, Floor Scrubbers and Vacuum Sweepers
  • I have been predicting simple robotic floor sweepers, floor scrubbers and lawn mowers for years, and for years, it hasn’t happened. The bottleneck may lie in the need for computer vision, which isn't yet cheap enough for these portable applications. Sooner or later, they're bound to reach the marketplace. The biggest problem is that of knowing where the sweeper or lawn mower is located at any given moment. The next biggest problem is coping with unexpected—e.g., your son left his bicycle in the middle of the yard. They will probably be used first in commercial applications, where labor costs are involved.

    •     Robotic vacuum sweepers are probably the easiest to develop, since vacuum sweepers operate in small enclosed spaces and since they don’t pose the hazards traditionally associated with lawn mowers. Electrolux is recently test-marketing an $800 floor sweeper the navigates a room, using ultrasound to "see" the room.
        Dr. Hans Moravec has estimated that computational speeds of the order of 10 terops (10,000,000,000,000 operations per second) will be required to match the computing power of the human brain. . However, with speech synthesis, and speech, handwriting and optical character recognition, we are already encroaching upon rote-mechanical, higher-level human functions.
        Of all man’s inventions, robotics and artificial intelligence may well be the most profound. It’s happening as we watch, in the form of speech recognition, facial recognition, speech synthesis, natural language responses, and smarter and smarter data base search engines. We’ve come a long way in just the last few years. I'm afraid to set a timetable. The Matrox Genesis Six-Board Imaging System currently delivers more than 100,000,000,000 operations per second. It seems reasonable to suppose that by 2007, 10 years from now, its successor may provide the requisite 10 terops.     (5) Better Communications
  • In contrast to computer technology, which has progressed at warp speed, telephone technology, as perceived by the user, isn’t much farther ahead today than it was when Alexander Graham Bell invented it in 1876. Technically by now, telephone conversations could take place in CD-quality sound but instead, they sound just as I remember them sounding in 1935. Our consumer-discernible telephone innovations have probably occurred only because of the Carterphone Decision that forced the Bell System to allow foreign devices to be attached to telephone lines. ISDN was touted as the next step in communications bandwidth, but it has been priced beyond the point where most consumers want to buy it, especially since it can’t also be used for voice. The regional Bell telephone companies have a monopoly on telephone service and they charge about $550 per month per megabaud for bandwidth. Why should they reduce the price of bandwidth when they can hold the customer for ransom? As one author puts it they are moving into field trials at the speed of "a snail on valium". The author points out that the regional Bell operating companies are more adept at fighting competitive deregulation in the courtroom than at improved delivery of services. After two years of field trials, none of the regional Bells have provided a single commercial rollout! There is now a lot of hype in the electronics news media about "splitterless ADSL (Asymmetric Digital Subscriber Line)" services. In the meantime, technological wizardry has given us 56K modems. Now, modems are under development that can multiplex data simultaneously over two telephone lines to double our bandwidths (to 67.2 kilobits per second upstream and 112 kilobits per second downstream) at only twice the cost of a single telephone line. Also, better data compression and local-cache techniques are appearing that may expedite data transmission over existing telephone circuits.

    • Microsoft and Intel have bet their money—several billion dollars of it—on cable-based data services.
        Cable companies have far greater incentives than telephone companies to move into this (for them) new area of data services. Cable services such as @Home are typically running $40 to $50 a month, including the Internet Service Provider fee, and are typically providing 1.5 Mb/sec. download speeds and 0.3 Mb/sec. upload data rates. This translates into about $13 per month per megabaud downstream and about $65 per month per megabaud upstream. Cable data residential service accounts, now numbering perhaps 100,000 nationwide, are expected to reach 1,000,000 subscribers by the middle of 1999, and 7,000,000 subscribers by 2002. Telephone residential xDSL accounts currently number about 1,000 and are expected to rise to perhaps 150,000 by 2002. The telephone companies will probably concentrate upon their less cost-sensitive business customers who may not have access to cable data services. In other words, they will continue to gouge their business customers.
        Cable Alabama currently provides cable data services in the Huntsville area. At $99 a month, it is too expensive for most residential users but may be suitable for small office/home office users, or even larger businesses.
        Satellite and wireless services are inherently expensive, and are dependent upon your telephone line for upstream data transmission.
        It has recently been suggested that data might be transmitted into the home over electrical power wires.
        If competition can ever gets a toehold in the telecommunications business, the cost of bandwidth may drop dramatically, but in the meantime, we are stuck with whatever the Bell Systems monopolies foist upon us, at least until Comcast brings its @Home service to Huntsville.     (6) Net Video-Casting
  • Video broadcasting and entertainment over the Internet is currently available, but TV-caliber broadcasting awaits much-wider bandwidth communications. For most of us, that may not occur until the next millenium. However, CDs could bring video entertainment to us before that time (which, in a sense, they’re already doing).
    •  (7) Video-Telephony
  • Like video broadcasting, video-telephony requires higher bandwidth than is currently available to most of us. Also, it requires video equipment at both ends of the telephone line. It will probably be the next millenium before videoconferencing is in widespread use (3 to 5 years).
    •  (8) Virtual Reality
  • Virtual reality could become the hottest entertainment medium in history. In a sense, it already is, in the form of computer games. It is extremely computationally demanding, and can profit from orders-of-magnitude improvements in computing speeds. It would also profit from 3-D displays, tactile feedback, and other enhancements. Ultimately, it will probably supplant conventional TV, including VCRs.

    •     One of the biggest obstacles facing virtual reality is the orders-of-magnitude speed reduction that arises when a graphics program is written to run under a Windows 95 or Macintosh operating environment. Most computer games have been written in DOS to achieve the speeds of which computers are capable. However, Microsoft has recently released a software development kit called WinG SDK that permits high-speed graphics under Windows 95.
        I think that a promising application for virtual reality lies in the area of online merchandising. Apple Computing has just introduced a QuickTimeVR kit for easy "stitching-together" of 4p-steradian photographs of objects taken from all angles, so that you can rotate virtual objects on your screen. You can also zoom in and out to get a better look at them. Such virtual objects would also seem to be ideal candidates for 3-D displays so that we could get a good look at that pretty basket that we might want to order online. This is an area where tactile feedback might also be valuable. This probably won’t happen next year but it would appear to be a good candidate for a marketing study. There’s probably a lot of money to be made by providing such services for online marketing. (The Mars Rover analysts are shown using red/green 3-D displays to get a feel for the Martian terrain around Sojourner.)     (9) Three-D Displays
  • Like robotic sweepers, I have been predicting 3-D displays for years, and like robotic sweepers, they haven't appeared yet. Three-D equipment is inexpensively available but for some reason, it hasn't yet caught on. I believe that virtual reality displays are likely candidates for 3-D imagery.
    •  (10) Toys That Converse with Children
  • Conversational dolls and toys are dependent upon hardware prices dropping low enough to fit in toys. Memory will be the key to this. Once RAM prices fall to, perhaps, $0.25 a megabyte--which should happen by 2003--it should become feasible to incorporate a general-purpose voice-recognition and voice-synthesis engine in a toy. It may even be possible earlier than that. By 2005, this level of embedded computing capacity should pose no problem. In the meantime, the software to support such interactions with children should steadily improve. Who knows what toys will be like in 20 years?
    •  (11) Autonomous Vehicles
  • In concert with federal and state government agencies, a consortium comprised of GM, Toyota and Honda is testing experimental autonomous vehicles on a 7.6-mile stretch of Interstate 15 near San Diego. GM and Honda are using small magnets embedded in the road to steer their cars. Toyota is also using this technique, along with a computer-vision centerline follower to steer its cars. The Toyota approach will work on any road with a reasonably visible centerline. It is anticipated that autonomous vehicles will first be used in HOV (High Occupancy Vehicle) lanes. Autonomous vehicles entering these lanes will be electronically checked for operational integrity before being allowed to drive autonomously. The timetable for widespread introduction of autonomous vehicles is of the order of 20 years.

    •     I suspect that this timetable will be influenced by how rapidly autonomous highways are adopted by other countries. The real payoff will come through trucks and other commercial vehicles. Also, emotions will play a large part in this timetable. The first time an autonomous vehicle has an accident, and especially, the first time an autonomous vehicle causes a death, the media will have a field day. (The safest, most practical course would probably be to build or set aside interstates exclusively for use by autonomous trucks. That way, if there were accidents or problems, no one could be hurt.)     (12) Printers, Scanners, Color Fax Machines, and Color Copiers
  • High quality, low cost multipurpose color printer/scanner/fax/copiers have opened the door to what is, perhaps, the last chapter in the home use of these devices. Improvements that might still be made in them might take the form of lower prices, higher fax resolution and color depth, water-resistant inks, photo-realistic color, and special effects kits (such as gold and silver printing, plain-paper printing, or six-color printing). However, in a few more years, they will probably set the standard for new hard copy devices for computers.
    •  (13) Removable Disk Drives
  • With disk capacities climbing into the gigabyte range, something must be chosen to replace the floppy drive, but at this time, it isn’t clear to me what that will be. So far, Zip, Jaz and tape drives have gotten the nod for hard disk backup. However, at $100 a gigabyte, the removable media are as expensive as an external disk drive. 120-MB floppy disk drives are available but they cost as much as Zip drives, as do their floppy media. The most practical removable disk medium would seem to be a write-once or rewritable CD drive. Writable CD disks are relatively cheap and store 650 megabytes of data. Every computer has a CD drive. However, greedy squabbles have characterized the DVD market and it’s debatable when DVDs will finally become standardized and popular. It’s hard to forecast what might happen here.
    •  (14) Video/Digital Cameras
  • Prices of bottom-of-the-line color video cameras are expected to drop as low as $30 by the end of 1998, with respectable color cameras available for about $60. (First-generation QuickCam Color Cameras dropped in price from $230 in the spring of 1996 to $100 by the end of 1997.) However, competing standards, lack of proper Internet hubs, severe bandwidth limitations, and the fact that the people with whom we want to videoconference aren’t equipped for video has kept videoconferencing from widespread use. Perhaps in the 1999-2000 time frame, videoconferencing will finally make it into prime time. "Talking head" video conferencing is said to be tolerably good over voice-grade phone lines. One of the problems is that of simultaneously carrying both audio and video over voice-grade lines with such a narrow bandwidth. If we ever get wide-band line service at voice-grade prices, videoconferencing may take off.

    •     Digital cameras have been expensive, with low resolution. In addition, they need to be viewed on a computer and/or printed out on a color printer. As previously mentioned, color printers are becoming photo-print capable and cheap. Also, digital camera price/performance ratios are improving rapidly. Umax has just introduced a $400 camera that delivers 1,024 X 768 resolution. It will probably be several years yet before digital cameras begin to crowd out conventional film cameras, but with the advent of low-cost, high-quality color printers, it will probably happen. Simple semiconductor technology extrapolations would suggest that 2,048 X 1,536-pixel $400 digital cameras might be available by 2001, and 4,096 X 3,072-pixel $400 cameras might appear in 2004. The 2,048 X 1,536 camera in 2001 would provide 7" by 5" 300-dot-per-inch prints and might render such a camera quite attractive. A Year-2000 1,024 X 1,280 $200 camera would yield 4" X 3" prints and might sell well. In summary, digital cameras are coming but are not quite here yet. (There is also the problem of rapid technological obsolescence.)     And Looking Still Farther Ahead…
        As previously mentioned, the Semiconductor Industry Association has an unpublished technology roadmap through 2022. As currently envisioned, circuit design rules will diminish to 0.05 m by 2012.
        There is evidence that materials still exhibit bulk properties down to 0.03 m. If that can be stretched to 0.02 m‘s, then conventional semiconductor progress might continue to follow Moore’s Law through 2022, with computer price/performance ratios halving every 18 months.
        If so, then by 2022, your bargain-sale desktop computer will be equipped with 2 to 4 terabytes (2 to 4 million megabytes) of RAM and will run at a speed of 20 to 30 trillion operations per second or about 100,000 times that of a 166-MHz MMx Pentium. This hardware capacity should be sufficient to support human-level intelligence at a readily affordable cost. However, as mentioned in item 4 above, near-human intelligence may be reached with the aid of specialized digital signal processors well before 2022. The biggest obstacle here is probably that of software.
        Even if circuit shrinkage were to stop abruptly in 2012, there should still be about six years worth of price/performance improvements for RAM, as 256-gigabit RAM chips go into volume production. This should reduce RAM prices to, perhaps, $200/terabyte by 2018. One more round of shrinkage, (or die-size expansion) to permit 1 terabit RAM chips, might reduce RAM prices to $50/terabyte by 2022. Of course, at some future time, in all likelihood, the rate of semiconductor shrinkage will probably slow from its current hectic pace but will probably not stop altogether. Or, while transitioning to new technological approaches, the rate of progress may slow for a year or two and then speed up again.
        The use of multiple processors offers an alternative approach to higher computer speeds when uniprocessor speeds finally peak out. The ability to pack billions of transistors on a chip should make it feasible to mount several microprocessors on one die. And of course, several such CPU chips, each containing several microprocessors, are possible. Costs would be somewhat higher than for single chip systems but not distressingly so. Furthermore MMx instructions can run several times as fast as the main desktop processor for those types of calculations at which they excel.
        How far can these ponies run? If by 2025, we reach the $20/terabyte-of-RAM, 40¢/terabyte-of-disk, hundred-terops speed range, then from a hardware standpoint, we ought to be able, easily and cheaply, to simulate human-class intelligence. This would probably be done using mass-produced specialized hardware and software. However, this may actually be achievable by the 2008-2010 time period, using special purpose accelerators. The real problem is going to be software.

    Appendices A, B, and C will be transmitted later under separate cover.