The next 50 years: More Change Than Anyone Can Imagine
Gordon Bell
gbell@microsoft.com

In 1947, the big idea (perhaps of all time) was the stored program computer that was soon to operate. In the same year, the transistor, or second and equally big idea, was discovered and by the mid 1960s a way of fabricating and interconnecting transistors on silicon substrates was invented and in use.

The development of the microprocessor in 1971 insured that the evolution would continue in a very focused fashion. The next 15-25 years looks equally bright. The only form of intelligence more easily, cheaply, and rapidly fabricated is the human brain, estimated to have a processing power of around 1000 million million ops/s, (one Petaops) with a memory of 10 Terabytes (Cohrane 1996).

For five decades, hardware has stimulated the evolution of computer platforms of various performance, size, cost, form, and applications from watches and pacemakers to mainframes. It is safe to predict the 2047 computers will be at least 100,000 times more powerful. If hardware continues to evolve at the annual factor of 1.60 rate we know as Moore's Law, (Moore, 1996) then computers that are 10 billion times more powerful will exist! Magnetic storage density and fiber optic data transmission rates, have evolved at the 60% rate (a doubling every 18 months, or 100 times per decade), too. It is also likely that since improvements in algorithms and methods often occur at the same rate as hardware, any future goal is likely to be reached in half the time one will predict based on hardware alone. It is difficult to speculate that the homely computer built as a simple processor-memory structure will take on a very different look, but rather continue on an evolutionary path of only slightly more parallelism of instruction execution. For the last decade, real application performance (RAP) of microprocessors has diverged from peak announced performance (PAP) that follows Moore's Law. This trend will continue!

Figure 1 shows the past hardware evolution and a 50 year extrapolation into the future. The next 15 years, based on semiconductor progress is likely.

What forms does the future computer take?

All intellectual property and everything bitable will be in Cyberspace.

With Cyberspace, the speed limit is our ability to find new places.

Bitability comes from the hardware and software interfaces (I/O) that the computer has acquired, created, or evolved to allow it to communicate with people and the physical world. We eventually expect speech, video, and gesture interfaces followed by having computers that anticipate. Surely, we can expect a "do what I say metaphor" within a decade since it has been a dream so long. Direct body interfaces are increasingly important, including touch, direct nerve stimulus, and artificial organs, eyes, ears, and limbs. It is difficult to predict that computers will interface by taste and smell. For achieving mobility and navigation in the physical world that permit useful robots, the big inventions exist as demonstrations, with video recognition, GPS, laser sensing, single chip phased array radar, and sonar. They have to be evolved to be low cost components and to be fast enough. By 2047 we would expect to see useful robots in homes, commercial areas, and factories that do not require extensive training.

New computer classes based on price will continue to be determined by applications and their resulting markets together with three factors: hardware platform technology (e.g. semiconductors, magnetics, and displays) , hardware- software interfaces, to connect with the physical world including people; and network infrastructures (e.g. Internet, and eventually, home and body area networks).

My theory of computer class formation based solely on using lower cost components and different forms of use to stimulate new structures accounted for the emergence of minicomputers (1970s), workstations and personal computers (1980s), and personal organizers. The world wide web using Internet has stimulated other computer classes to emerge, including: Network Computers, telecomputers, and television computers that are combined with phones and television sets, respectively. This basic theory also accounts for the emergence of embedded and low cost game computers using world-wide consumer distribution networks. Mobility via a radio networks opens up more future possibilities that are not just adaptations of cellular phones.

Within a few years, scalable computing using an arbitrary number of commodity priced computers and commodity high speed networks are likely to operate as one and replace traditional computers, i.e. servers of all types! We call this approach to computing based on upsizing, SNAP for Scalable Network and Platforms (Gray, 1996). The underlying parallelism is a challenge that has escaped computer science for decades.

As communication instruments, computers enable the substitution of time and place of work, creating a flat, equal access world (CNRI, 1996). After nearly thirty years of the Internet, people-to people communication via email and chat remain the top applications. Is telepresence for work, learning, and entertainment the long-term "killer app"?

Can these be built in this short time? Or, even will the computer interface with humans biologically, rather than the superficial, mechanical way they do now?. More likely, nearly zero cost, communicating computers will just be everywhere embedded in everything from phones and light switches to all-seeing, all changing pictures. They'll be the eyes and ears for the blind and deaf, and eventually "drive" vehicles. We need to be fully connected anywhere at all time. The big idea is fiber optical cable that also evolves to carry more bits per second each year at1.6 times per year. Perhaps an equally big idea is in the making, the high speed digital subscriber link a.k.a. "the last mile" that permits high speed data to go to the home via the world's trillion dollars of installed copper connections. In parallel, radio links enable anywhere computing. Body and Home Area Networks are also part of the network story that need to be invented.

Table 1 New computer classes and the enabling components.
Generation Platform (logic, memories, O/S) User Interface Network infrastructure
The beginning (direct & batch use) vacuum tube, transistor,
core, drum & mag tape
card, paper tape none originally... computer was self contained
Interactive timesharing via commands integrated circuit,
disk, multiprogramming
glass teletype & glass keypunch, command language POTS using modem, and proprietary nets using WAN
Distributed PCs and workstations The Microprocessor PCs & workstations, floppy, disk, dist'd O/S WIMP (windows, icons, mouse, pull-down menus) WAN, LAN
world-wide web Evolutionary PCs and workstations, servers everywhere, Web O/S Browser fiber optics backbone, www, http
SNAP (Scalable Network & Platforms) PC uni- or multi-processor commodity platform server provisioning SAN (System Area Network) for clusters
One Dial tone: phone, videophone, tv, and data Network Computer, telecomputer,
tv computer
telephone, videophone, television, xSDL for POTS, cable, fiber (longer term); home area nets
Do what I say embedded in PCs, hand held devices, phone, PDA, speech, common sense Body Area Nets. IR and radio LANs for Network access
Anticipatory by observing user needs room monitoring, gesture vision, gesture control, common sense home area nets
Robots no special radar, sonar, vision, mobility, arms, hands IR and radio LAN
Ubiquity embedded $1-100 devices that interoperate computer to computer control home area and body area networks

References

Moore, Gordon "Gigabits and Gigabucks" University Video Corporation Distinguished Lecture, 1996.

Gray, J., Scalable Servers, http://research.microsoft.com/labs/barc/

CNRI, "Vision of the NII: Ten Scenarios", Reston, VA, 1996. See also papers on future networking. http://www.cnri.reston.va.us/

Cochrane, Many papers on the future of computers , including their use for caring for the aged.

Figure 1. Evolution of processing rate (in Mips), primary and secondary memory size (Bytes), and data rate (bits/sec) versus time