ith today's multitiered, over- substituted for that of traditional W lapping set of programmable minicomputers, these suppliers find classes, where and how themselves in a situation similar to computing can be done and how BUNCH'S dilemma. For example, much it will cost can vary con- SEL and Prime, minicomputer siderably. Computing costs can be manufacturers, have marketing/ anywhere from $100 to $10 million distribution agreements with Con- (Figure 1). In addition, computing vergent Technology, but mini com- devices can include electronic panies must compete with systems typewriters with built-in communica- built from high-performance, tion capability, further increasing the commodity-oriented, 32-bit MOS- choices to be made and the complex- based microprocessors-processors ity of the information processing that provide the same performance market. as the traditional TTL-based pro- What is happening to mini and cessors at a small fraction of the cost. Market pressures have mainframe companies as the micro In short, the forecast could be forced marginal continues to pervade the industry? gloomy for mini companies. Just as One thing is that several traditional the mainframe companies were un- computer firms out of mainframe suppliers, Burroughs, able to respond to the mini, the mini business. Looking to Univac, NCR, CDC, and companies will have difficulty mov- compatibles, wary (or BUNCH, for brevity's sake), are ing to meet the micro challenge customers are helping experiencing a declining market create de facto share as mainframe customers select standards. IBM-compatible hardware as a stan- Table 1. Minicomputer technology cir- dard and turn to other forms of com- ca 1970. puting. Fujitsu, Hitachi, Mitsubishi, and NEC supply commodity main- BASIC MINICOMPUTER frames, which are distributed COMPONENT SYSTEM through Amdahl, National, Univac, INDUSTRIES COMPANIES and Honeywell. Power Supplies Optional Packaging Essential The microprocessor-based systems Core Memory Optional are the newest alternative for distrib- Semiconductors CPU and Memories uted computation. New companies (MSI) are forming to develop these pro- Disks and Tapes Peripheral ducts; Burroughs and NCR have dis- Controllers Terminals - tribution agreements with new Operating Systems microprocessor suppliers such as Languages Convergent Technology. As micro- Applications processor technology continues to be System Integration

COMPUTER because of the large installed bases, with other companies. Established introduced the Series 1. Several com- proprietary standards, and large small and com- panies, Floating Point Systems, for functional organizations. panies such as Scientific Data Sys- one, were started up to build special tems and CDC attempted to develop signal- and image-processing "niche a line of minis, and other electronics- to supply high-availability and The minicomputer generation related companies looked at the op- cluster-expandable minicomputer portunity to enter the computer systems. In the beginning of the minicom- business. We can make several conclusions puter industry, a product took two No significant minicomputer com- from the data on the minicomputer years to reach the market. This panies were established after 1972. In companies: period began with the start of hard- the late 1970's, IBM decided that ware design and went through writ- distributed departmental computing, Seven successful minicomputer ing an assembler, a minioperating using multichannel distribution companies-or eight percent of system, and utility routines for the (OEM/end user), was not a fad and all tries-survived to enter and sophisticated users. A relatively wide range of technology (Table 1) was re- quired to design logic, core mem- ories, and power supplies; to inter- face peripherals and do packaging; and to write system software, such as operating systems, compilers, assem- blers, and all types of applications software such as message switching. Clearly, this industry was high-tech. The early minicomputer, charac- terized by a 16-bit word length and 4K-word memory, sold for about $10,000. It was small and could be embedded in larger systems (for ex- ample, electronic circuit testers and machine tools); it could be evolved to large system configurations; and it was used for departmental timeshar- ing. Applications varied from factory control to laboratory collection and data analysis, and communications to computing in the office and small business. The original equipment manufacturer, or OEM, concept was established so that hardware and software and software-only applica- tions could be designed and mar- keted in two applications, thereby in- creasing the market for what was basically a general-purpose com- puter. Many more markets were created than could be reached by a single organization with a limited view of applications. From 1968 to 1972, about 100 minicomputer efforts were started by four different kinds of organizations (see box on following page). At least Figure 1. 1984 system price versus machine class. The dots on the ends of 50 new companies were formed by the lines signify the uncertainty of price range. Because these classes are relatively new, prices are changing rapidly. Also the class has a broad defini- individuals who came from estab- tion; that is, a number of products of varying complexity can go by the same lished companies or research labora- name. Products within a class can be anything from boards to complete tories. Some of these later merged systems.

October 1984 15 defend themselves in the micro- processor market. Another 16 companies suc- ceeded to a lesser degree and still exist in either diminished or niche segments of the market. Of all organizations, 23 (25%) were successful. While virtually all companies built working , 75 percent did not build organizations with any longevity for a variety of reasons, including failure in engineering, failure in market- ing, faulty manufacturing, or insufficient product depth or breadth. Only two of 50 (4%) start-ups succeeded and remained inde- pendent, although nine of 50 (18%) continued in some fashion. For start-ups, merging increased the chance of survival; four of 60 (7%) could be considered winners. The probability of a successful merger was 50-50. An organization that is part of a larger body in some other busi- ness is pretty likely to fail; only HP-one of 23-really made it. A start-up within a large ex- isting company may as well be a stand-alone start-up. Companies selling in a different market or price band were un- able to make the transition. Only DEC made it, but we can argue that DEC was already in the mini business and simply maintained its market when everyone else started making minis. IBM eventually started making traditional minis in the late 1970's with the Series 1 and began claiming a significant market share. The System 3 (cir- ca 1972) was the most successful "business minicomputer." Companies that differentiated their products by using special- ized hardware and software were prone to failure. Vendors

COMPUTER that made special computers for modern parts could be used and the of complex software. By 1975, an application such as commun- experience of others could be taken several different operating systems ications or testing (real-time into account. The simplicity of the were available for the various market control) piways failed to make DG product allowed rapid under- segments. successful minis and often failed standing, production, and distribu- DEC converted the PDP-11 to a or fell behind in developing tion, especially to OEMs. The OEM multichip set relatively early and their main product. Specialized form of distribution is particularly entered the board market to compete hardware limited the market in- suited to start-up companies because with microprocessors to some stead of broadening it; although a product is not used in any volume degree. Until just recently, it led the specialized software could until one to two years after the first 16-bit micro market, but now chip- sometimes leverage sales, it was shipment. based micros are commodity parts, typically inadequate when used Prime, another successful start-up and the assembly of personal com- with limited hardware for a minicomputer company, arose under puters has become trivial. DEC single market. different circumstances. Before the failed to license the PDP-11 chips or In the mini generation having a company was established, Bill make them available for broad use, high-performance, low-cost, Poduska, its founder, had built the including the transition to personal general-purpose minicomputer breadboard of a large, virtual mem- computers, so unfortunately the suitable for broad application ory in a NASA laboratory. Prime PDP-11 today is merely another in- ensured getting the largest mar- was thus able to introduce the first of teresting machine that failed to make ket share. DEC, for example, the "32-bit (address) minis" in 1973. the generation transition. had a variety of operating sys- With this new technology, programs DEC introduced Vax-11, a 32-bit tems aimed at the real-time, such as CAD could be run. DEC mini, about six years after Prime in- single user (which laid the foun- didn't provide a large, virtual troduced its model, but at a time dation for the CP/M operating memory capability until 1978 when it when physical memories were large system for personal computers) introduced Vax. enough to support virtual memories and provided communications, The start-up of both DG and and provide optimum cost and per- real-time control, and timeshar- Prime were characterized by superb formance. Because it had much ing. The real-time system was marketing followed by the establish- larger manufacturing and marketing ultimately extended for transac- ment of a large organization to build divisions, DEC quickly regainid the tion processing. Minis became and service in accordance with de- market it had lost to smaller especially useful for business mand. manufacturers including Prime. applications because they were designed for high throughput. DEC-a steady force in the mini IBM-a consistent winner. IBM Although business computers market. After several false starts, always responds to mainline com- weren't useful for real time, DEC was able to compete with DG puting styles and needs, even though minis designed for real time and other start-ups because of its it sometimes enters the market late; were very good for business and momentum in three other basically for example, it didn't realize early on timesharing uses. mini product lines. Thus, its fun- that the minicomputer had broad damental business from its inception market appeal. in 1957 was small computers, and IBM sometimes innovates with DG and prime-the first market while it produced the first large- radical new technology such as the successes. The initial Data General scale timesharing computer in 1966, disk, chain printer, and Fortran, but and Prime products were unique and it also produced the first mini, the often follows pioneers in computing had a relatively long time to find a PDP-8 in 1%5. styles as evidenced by its develop- place before the established leader, With the onslaught of minicom- ment of the minicomputer, timeshar- DEC, reacted to the threat. DG was puter start-ups (including DG, which ing, the PC, local area networks, and established by engineers who had you will recall was formed by former home computers. built successful products at DEC (in DEC engineers in 1%8), DEC finally Some of its low-cost computers contrast to many start-ups that had responded with a competitive ldbit admittedly were nearly minis: the little or no experience in designing minicomputer, the PDP-11, in 1970. 1130 (1965) for technical computing, products). DG had a simple-to- The 11, which was comparatively the 1800 (1%6) for real-time and pro- build, yet modem, 16-bit minicom- complex, sold as a premium product cess control, and the System 3 (1971) puter based on integrated circuits and allowed DEC to quickly regain for business. In fact, while the that enabled it to be priced below all the market. With the PDP-1 1's minicomputer was forming, IBM was existing products despite its late en- Unibus, interconnection of OEM preoccupied with introducing the trance into the market. In fact, the products was easy, and extensive 360. However, we should remember late entry was a benefit, since more hardware facilitated the construction that the antitrust suit against IBM

October 1984 17 started in January 1969 and may ac- tionary change is required. Tech- have been observed in a single, large count for its lack of aggressiveness nology transition, which typifies the monolithic functional organization. during this time. generations, requires much change The industry formation process, IBM waited until PCs were estab- including a new computer, a new expressed in a style similar to Pascal lished before it entered the market market, and a new way of com- language dialect, is shown below. and established the standard. Now, puting. Since existing companies are procedure Entrepreneur-Venture-Cycle only two years after entering the unlikely to address a new market, begin market, it has the largest market new companies are required. jJi& Frustration > Reward {Push share. Today, IBM is tackling the from Old-co) and Greed > ear (hll to New difficult problems presented by company) do home computing. Thus, because of The microprocessorgeneration -begin its size, IBM can dominate any (and get (PC, spreadsheet); perhaps all) market segments of in- The micro-based information- IF System-Company then formation processing in just a few processing industry is composed of write (Beat-Vax-Plan)- years. thousands of independent, entrepre- ELSE write (Plan)- If we look at computing in the neurial-oriented companies that are New-Company simplest way-that is, in terms of stratified by levels of integration and get (Venture-capital); substituting alternative price and segmented by product 'function- {from Old-Venture-Co) performance levels-we can say that whether microprocessor, memory, exit {job); a low cost means more people can floppy, monitor, or keyboard- start (New-Company); decide to buy a product whether they get (Vax, developmenttools); within a level. build (product); sell (product); are small company presidents or The first computer companies sell (New-Company); department heads in a large com- built the whole system from circuits {@ 100 x sales) pany. The cost per user, then, deter- to tape drives through end-user ap- venturefunds := Co.-Sale mines the product's attractiveness plications in a totally vertically inte- start (New-Venture-Co.); end when weighed against other forms of - grated fashion. A stratified industry, -end computation. By both measures, on the other hand, is a set of in- IBM missed the minicomputer mar- dustries within an industry, each The "push and pull" concept. ket until it introduced the Series l in building on successive product The WHILE clause in the above (the 1977. layers. Each company designs and start-up) is evoked by two condi- In short, IBM will consistently builds only a single product within tions: the "push" of an old com- win, not only because of its size, but each level. Systems companies then pany and the "pull" of a new com- also because it aggressively views all integrate collections of the seg- pany or product idea. Throughout forms of computing and possibly mented products to produce a sys- each generation, we've seen the communication as part of its market. tem for final use. "push." Bill Noms led a group Three factors have caused this in- (including Seymour ) from HP-the only established com- dustry structure: (1) entrepreneurial Remington Rand's Minneapolis pany to succeed. Hewlett-Packard energy released by venture capital; group (originally Engineering purchased a small start-up called (2) standards,' which become con- Research Associates) to form CDC Dyrnec to enter the minicomputer straints for the products and create in 1957. Cray left CDC in the early business, and thus might be con- product divisions, or strata; and (3) 1970's to form Cray Research. Gene sidered a merger even though it in- the establishment of clearly defined Amdahl could not build high- tegrated the product into its organi- target product segments-so many in performance 360's within the IBM zation right from the start. HP's fun- fact that we are forced to ask "What environment, so he left to form Am- damental business was to produce in- part of the industry is high-tech?" dahl Corporation. Later, he left Am- formation from instrumentation dahl to form Trilogy for similar equipment, and it regarded com- Entrepreneurial energy. Com- reasons. Bill Poduska, who founded puting as fundamental. For most panies form in an entrepreneurial Prime in the early 1970's, came from companies outside the computer fashion and are able to participate in a NASA laboratory where he had field, computers were too much of a every level of integration in a single built a prototype of a minicomputer diversion from what they understood product or through the integration with a virtual memory. Later, he left and could manage. of products into a complete system. Prime to found Apollo Corporation The success of HP alone only The amount of energy released to and build clustered workstations. underlines a concept that usually build products through entrepre- Bob Noyce left the Schockley Tran- holds: Leaders in a market segment neurial self-determinism is truly in- sistor Company to form Fairchild of an industry usually remain credible; improvements in produc- (where he was a major inventor of leaders, unless too much evolu- tivity by a factor of several hundred the IC) and then left Fairchild with

October 1984 Grove and Moore to form Intel to financial projection, character- levels of integration form the in- develop the first MOS memories and ized by the desire for a practical dustrial strata, the bottom four being microprocessors. By most accounts, strategy that would yield high hardware and the top four being all these transitions were made with yet realizable returns and that software and applications. at least 50 percent push from the could be used as an operational Discipline and profession-spe- parent company. "yardstick." cific vertical applications. Two business plans, separated by Standards. Formal standards CAD for logic design and cir- the IF clause in the entrepreneur- developed by international standards cuit design and small business venture capital cycle, are (1) a com- groups established many of the stan- accounting. ponent plan to enter and address one dards (constraints) observed by Generic application. Word segment of the market, such as a new today's designers. These restrictions processing, electronic mail, spreadsheet package, and (2) a plan have gradually caused industrial spreadsheets. to build a computing system that will layers to form, which have clearly Third-generation program- win against Vax or some part of the defined limits. The following eight ming languages and databases. IBM PC market. Money is secured from one or more venture capital companies. The founders leave their jobs and start the New-Company in almost a single step. In some instances, "seed" financing is acquired whereby founders actually leave their jobs before the first business plan for the new company is written. Building and selling the company. The company proceeds to get a Vax for use as a development computer. They develop and sell a product. After the first profitable quarter the company goes public and the valua- tion is placed at multiples of up to 100 times the annualized sales of the company. (A multiple of slightly over one is not uncommon for mature but still profitable com- panies.) With the funds from the public sale, New-Venture-Co. can be formed to invest in new high-tech companies.

The start-up and two alternatives. A PC running Lotus 1-2-3 is required to write the plan and address the financial aspects (i.e., profit and loss and balance sheet). Poduska's ele- ments in a successful business plan, which must be less than 10 pages, in- clude2 summary-one page; market brief, a synopsis of who will buy and why; product brief, the what, why, and how of product building; people, the rule being use only Grade A, experienced people; and COMPUTER I + Fortran, Basic + Pascal + bus, Multibus Multibus I1 (proprietary, corporate process (evolution). and VME. standards. . . require formali- Operating system. Base sys- Standard chip. Micros, micro zation to realize a silicon- tems, communication gate- peripherals and memories/evo- foundry-based industry). ways, databases/integrated lution of Intel and Motorola Signal transmission, physical envi-

Basic + CP/M + MS/DOS + architectures synchronized to ronment, communications links, and Unix (evolution). the evolution of memory chip language standards have played a Electromechanical. Disks, sizes-8080 [S100](4K) + 280, key role in defining these strata. De monitors, power supplies, en- 6502 (16K) + 8086 [Multibus, facto standards by various manufac- closures/8" + 5" + 3"(?) flop- PC Bus] and 68000 [VME] turers, which provide the most im- py; 5" Winchester (evolution). (64K) + 286 [Multibus 111, portant standards, are micropro- Printed circuit board. Buses 68020 and NS32032 (256K). cessor architectures, buses, periph- synchronized to micro and Silicon wafer. Bipolar and erals, operating systems, and applica- memory intros/S100 + PC evolving CMOS technologies tion software file formats. Regret- tably, we often misunderstand and underestimate the importance of these and other standards.'.' Product segmentation. The num- ber of clear product segments in the industry is a major determinant of its present structure. To understand that structure, we need to isolate which products are worthy of the title "high tech." Advanced technol- ogy is characterized by significant in- vestment, highly skilled personnel who understand the technology, and often high project risk. Products evolve at a rapid rate and demonstrate continued performance and price improvements, together with innovative structures. The resulting products demand a premi- um. High-density semiconductor and magnetic recording products fit the definition, but most systems assembled from these components, such as IBM-compatible PCs, are clearly not high-tech because they are simply a system formed from high-tech components. The barriers for entering an end- user, OEM, or system-level business with a generic product are not very imposing (Table 2 shows the technol- ogy requirements), especially when they are compared with the complex- ity of the engineering needed to pro- duce early mainframes and minis (Table 1). A micro-based system company can be formed by a part- time president, someone with a PC and Lotus 1-2-3 to do the business plan, someone who can buy and assemble the various circuit boards into a Multibus backplane, a pro-

October 1984 21 Table 2. Microcomputer-based tech- computer class (see box on previous mance as their larger computer nology circa 1978. page). Clearly, many more issues ancestors. In addition, micro-based are involved in accounting for per- products can be interconnected in a BASIC MICROCOMPUTER vast array, forming a much larger COMPONENT SYSTEM formance, price and relative perfor- INDUSTRIES COMPANIES mance/price, including machine age; range than ever before. The most im- portant structure to emerge is the Power Supplies Optional hardwired versus microprogrammed Packaging Optional control and associated instruction local area network, because it per- Semiconductors - times; memory speed; Vax's cache mits the formation of a much larger, (micros, memory, performance (neither the Cray nor potentially single system. peripherals) the IBM PC uses a cache); floating CRTs and Terminals - point speed; degree of parallelism for Disks and Tapes - LAN-based computing. The infor- both vectors and scalars; the relative Board Options Optional mation processing structure within a (displays) goodness of the Fortran compilers; large organization is driven by newly Unix & Diagnostics Optional and actual use versus a single bench- emerging computer structures, com- Languages & Optional mark to typify a computer's work- Databases puting nodes, and local area net- load. LANs and Optional works, or communication links (Fig- Communication ure 2). The LAN is critical to com- Applications Optional System Integration Micro and mini computing puter evolution during the next few structures years, and the lack of standards is greatly impeding pr~gress.~ Hundreds more products can be The multiprogrammable operating grammer to buy and load a version built from the micro than can be systems introduced in the mid 1960's of Unix, and one or two helpers. built from the mini because of the allowed a machine to be shared by a The point I am making here is that micro's low cost, small size, and ease number of users, if each had a "vir- the single, most important measure of programming. Personal com- tual" computer (Figure 2a). Since of the high-tech portion of the micro puters, terminals, typewriters, and overloading is common in shared sys- industry is semiconductor improve- computing PABXs are all lower cost tems, users enjoyed having their ment. That is, semiconductor alternatives to larger computers that own personal computers when rea- technology mainly determines the provide relatively the same perfor- sonably powerful, reasonably cheap

Figure 2. Evolution from timeshared central computers to LAN-based clustered workstations and personal computers.

22 COMPUTER models were introduced in 1978 by Gateways must be done once work. (A WAN is usually not used as Apple and then in 1981 by IBM for a network or protocol in- a single system, but as a communica- (Figure 2b). PCs proliferated in large stead of for each system, there- tion network among several systems, organizations. The need to obtain by limiting the number of com- including LANs.) data from the shared computers munications protocols. The combination of micros, meant that programs had to be higher level performance, wide-scale developed that would allow PCs to The computing nodes. Figure 3 is a use, and higher reliability can be of- emulate dumb terminals. Increased taxonomy of common mini- and fered for the price of a mini or super- PC usage, coupled with greater ex- micro-based computer structures, mini. Complete new structures have pectation of response time, provided which illustrate the plethora of new emerged, including functional multi- a demand for increased shared com- computer structures made possible processors, symmetric multiproces- putation at minis and mainframes. by the micro. (For more details on sors for performance and high avail- Because users wanted access to specific structures, see the appendix ability, fault-tolerant computers, specialized and central data, the de- to this article, "Specific Microcom- and multicomputer clusters. In addi- mand for mainframes has resurged, puter and Minicomputer Structures." tion, microcomputers are combined and this trend is likely to continue These range from the simple PC to in fixed structures to provide high- until a fully distributed, LAN-based the LAN, omitting the wide-area net- performance, close-area-network system (Figure 2c) is built. Palo Alto Research Center invented the LAN-based cluster con- cept in the mid-1970's using Ether- net, the basis of IEEE 802.3, the LAN standard. For powerful work- stations such as the Xerox Star or Apollo Domain, the LAN must per- mit the sharing of files and intercom- munication of work. Functional services such as filing and printing of the shared system (Figure 2a) are de- composed into specialized "servers" (Figure 2c) and connected along a LAN. A LAN, then, must address several needs:

Large, shared systems must be "decomposed" for improved locality, lower cost, physical security, communication with a single resource, and incremental evolution. Personal computers or work- stations must be "aggregated" into a single system to share resources such as printers and files to intercommunicate. Networks of minis and main- frames, which have relied on poor wide-area, data communi- cations facilities for local com- munications, require high-speed intercommunication. The connection of minis and mainframes to terminals must Figure 3. Taxonomy of common mini- and micrebased computer structures. C = computer; P = processor; K = controller; Cluster = collection of C's acting be completely flexible, and in- as a single C-interprocessor communication times determine parallel process- cremental upgrades must be ing grain size; and function = arithmetic, array processor, signal processor, possible. communication (front end), database (back end), display, slmulatlon.

October 1984 computer clusters. If a method can dard operating system, complete large enough to serve as main be found to use a large number of product segmentation may occur to products very long. essentially zero-cost microprocessors eliminate vanity architectures at all Generic and unique software in various multiple-processor struc- levels of integration. applications like CAD that run tures to work on a single job stream, If minicomputer history is a good on a few standardized structures then micros can potentially compete indicator, fallout in the micro-based (PCs, workstations, and super- with all forms of computers in- industry will be even more legendary. micros) will fuel this generation. cluding mainframes. Fox4 has used For example, of the 100+ worksta- Truly unique structures like an array of 64 Intel 8086/8087-based tion companies, we can expect fewer home robots are rarely suffi- computers for particular theoretical than 10 to survive, let alone prosper. ciently protected by patents, physics calculations to show that this A similar statement can be made processes, or practice to avoid structure can approach supercom- about the PC market. The following becoming displaced by an estab- puter performance. criteria will determine success: lished supplier entering the Figure 4 illustrates the variation in market. Remember how quickly processor types for common com- Economy of scale in distribu- IBM became a dominant force puter types. Micros have followed tion and service is most impor- in the PC market? the traditional mini evolution and tant. are today microprogrammed with * Economy of scale in manufac- the exception of the MIPS chip at turing is critical for a few The applications challenge Stanford5 and the RISC chip at the focused products such as the University of California, Berkeley." PC but less important for larger Now that we have examined the Given the current speed of logic products. Here, systems inte- bewildering number of products and relative to memory, it is again time to gration costs dominate. For ex- services available, we need to look at return to direct (versus micropro- ample, the Japanese are likely ways to supply them. A number of grammed) execution of the instruc- to dominate the PC market in strategies are possible, from selling a tion set when performance is a con- much the same way they domi- purely general-purpose base system sideration. nate consumer electronics. to offering customized hardware and Time to market is far more im- software. In the latter case, however, portant than economy of scale the resulting function may scarcely The systems industry in engineering or manufactur- resemble a computer. Economy of ing. scale may occur in the widespread Virtually all microprocessor-based Since there are few techno- sales, distribution, installation, and systems supply a single information logical challenges in a start-up, service of hardware products. processing market. Micros allowed companies will form if they get An OEM approach usually re- the PC to form but also to attack venture capital; later entrants quires a product range, not just a the traditional minicomputer, the will be less successful. point product. An OEM customer high-availability mini, and possibly Specialized, or niche, products often requires service and always re- the mainframe. Now with the stan- are rarely "sacred" enough or quires high-level applications and field support. An end-user approach requires both a wide product range and complete sales/service. A new application software com- pany, such as one offering CAD or typesetting, that has to invent its own hardware system is likely either to become obsolete because of its hardware or to fall behind in its software development. The company is limited because investment has to be divided between its unique vanity hardware and its specialty, added- value software. In most cases, large hardware vendors, such as AT&T, DEC, and IBM, can surpass the small hardware/software supplier by using packaged software from the Figure 4. Taxonomy of common processor types. applications software industry.

24 COMPUTER Supplying the basic computer. Mentor Apollo (and the hardware tion, the software company doesn't Figure 5a shows the simplest form of supplier of the workstation) or it can usually market the range of products distribution for what is fundamental- be bought a la carte and integrated by of a mainline hardware supplier. Of- ly a computer sold with some gener- the customer. fering a total system, therefore, is al-purpose software. A base system Supplying applications software as likely to be less profitable-when would typically include generic soft- part of a system. Since the perceived measured by return on investment- ware such as languages, utilities, (and often actual) price of software is than offering software only, even editors, communications interfaces, low, a company marketing a software though the total revenue of the com- and database programs. The system product and wishing to enhance its pany would be much larger in the is built by a hardware manufacturer sales volume may buy hardware for former case. Furthermore, the sup- or system integrator; it is sold (S) resale as a complete system (Figure plier is cut off from the large number directly or through another distribu- 5c). In effect, a company potentially of distribution channels possible tion channel of some sort; and even- competes with the hardware's main when a basic software package is tually, the system is installed (I), the manufacturer by supplying a similar, tailored for operation on many dif- user is trained (T), and the system is but greatly enhanced product. While ferent base systems. Computer Vi- serviced (S). the gross sales are up, the costs can sion is an example of a company that easily outrun the sales, since the now buys products on an OEM basis Supplying the basic computer with once-software-only company must from Apple and IBM, integrates applications software. As users re- now support hardware too. In addi- them, and supplies them as turnkey quire more specialized applications for particular professional environ- ments, such as the computer-aided design of electrical circuits, various industries will supply these pro- grams, creating a product develop- ment and distribution structure (Fig- ures 5b, 5c, and 5d). The base-system manufacturer and an independent software in- dustry can coordinate the introduc- tion of applications programs into the distribution network (Figure 5b). Special software can be integrated with the base system by the hardware supplier, the application supplier, the distribution channel (store or systems installer), or the final user. A system manufacturer can ac- quire a variety of packages and transform what is a general-purpose system into a variety of special- purpose systems. The software sup- pliers are likely to be the best ob- tainable for the application selected because they have focused on the particular, vertical professional ap- plication, be it mechanical or elec- trical CAD, architectural drawing, office automation, or actuarial or statistical analysis. The software sup- pliers have the largest market because a program can be trans- formed to run on many different base systems. Mentor is an example of a CAD company with a flexible Figure 5. Alternative industry structures for supplying base, application and approach to systems integration. A hardware-embedded computer systems (S/I/T/S = sell, install, train, ser- total system can be purchased from vice).

October 1984 products. Computer Vision formerly manufactured its own base system. Supplying unique hardware and application programs. The tradi- tional approach of catering to OEMs, which DEC established with the minicomputer, is shown in Figure 5e. A company skilled in a particular technology-computed axial tomog- raphy is a good example-or in logic board testing can build a highly com- plex instrument. A computer may constitute up to half the cost of the system. Products of this nature are not basic, general-purpose com- puters, and as such, the customer will not require other software beyond the control of the device. A specialized field organization is re- quired to sell, install, and service the system and to train users. This sup- port is hardly possible with a conven- tional computer company.

A final word about applications. Applications that involved minicom- puters are likely to be a good history lesson. Companies that tried to backward integrate and build their own minicomputer, such as Cincin- nati Milling, failed in the market, often neglecting their mainline business. The applications system winners combined the use of a general-purpose mini with their ex- pertise in the application. Companies that use high-cost, vanity hardware or who distribute someone else's hardware will be at a disadvantage because the value of the product is completely in the software. New professional software ap- plication products will come from those in existing companies and in- stitutions such as universities who have expertise in particular problem domains. Applications industries will form and evolve through the strata model discussed earlier in "Stan- dards" to software-only companies that create the professional applica- tion (a form of "expert" system) and use standard systems supplied by hardware vendors such as Figure A-1. Common micro- and mini-computer structures. PC = central processor; IBM. Pio = i/o processor; K = control; Mp = primary memory; Mc = cache; Ms = secon- Thus, we have an opportunity not dary memory; and T = transduce (terminal). available in industry-to build

26 COMPUTER generic, basic hardware systems in a 2. w. D. Poduska, The Formation of Stratus provides a fault-tolerant sys- crowded field, resulting in an almost APO~~OCor~oratiofl, IEEE Engineer- tem (Figure A-If) that is completely ing Management Chapter, Dec. 12, unlimited set of professional applica- transparent to its software. Any hard- ,,,, ,,,,I/"-'. ware component can fail and the system tion products as experts encode their 3. C. G. Bell, "Standards Can Help will continue to operate without affect- "knowledge" into programs for Us," Computer, Vol. 17, No. 6, June ing the basic software. The single point machine interpretation and personal 1984, pp. 71-77. of failure is system and application soft- use. These will constitute the real ex- 4. C. G. Fox, "Concurrent Processing ware. Stratus systems require four pro- pert systems of the fifth generation, for Scientific Calculations," Proc. cessors and two memories to provide a as they run on evolutionary micro- Compcon Spring 84, IEEE-CS Press, single, effective processor. processor-based computers and Los Alamitos, Calif., pp. 70-73. Synapse N + I (Figure A-lg) uses a clusters of computers connected by 5. J. L. Hennessy et al., "The MIPS second bus for both performance and local area networks. Machine," Proc. lEEE Cornpcon redundancy in a true symmetric multi- Spring 82, pp. 2-7. processor version of the single bus ew technology, especially 6. D. A. Patterson and C. H. Sequin, system. By having all resources in a single pool, users can trade off performance VLSI, has provided powerful, "RISC-I: A Reduced Instruction Set N and reliability. Since work can be run on low-cost microprocessors and mem- VLSI Computer," Eighth Annual Syrnp. Computer Architecture, May any processor, load leveling is automatic. ory which, in turn, have acted as 1981, pp. 443-458. Tandem (Figure A-lh) pioneered high- standards and permitted a new in- availability computing when it intro- dustrial structure to emerge. The duced its multicomputer system in the structure, which is typical of a cot- mid-1970's using minicomputer technol- tage industry and is almost the anti- ogy. Sixteen computers are connected in thesis of a vertically integrated in- Appendix: Specific a cluster via a dual, high-speed message- dustry, is stratified by eight hardware and passing bus. Complete redundancy is and software levels of integration Microcomputer Structures provided, including computers, control and segmented by a vast array of units, and mass storage. Operating Figure A-1 illustrates a wide range of systems and applications are run in two component products. Companies are microcomputers, from the common, computers in a backup fashion. Informa- funded by a vast array of venture single-processor, "Unibus" structure tion is forwarded to the backup process capital companies formed from the (Figure A-la), to computer clusters for using the intercomputer bus. A key use profits of selling previous companies. high availability (Figure A-lb). Since of the Tandem structure is to permit in- The resulting products are integrated microprocessors require memory access cremental addition of performance. into an equally large array of system at a higher rate than the first DEC Since processes and files are assigned to products by traditional system sup- Unibus and Intel Multibus (2M and 4M specific processors, load balancing is less pliers, such as IBM; companies that bytes per second), the common adapta- dynamic than that in the multiprocessor. add value by distribution, service, tion is to provide a direct connection be- Several microprocessor versions of the and training; conventional, retail tween the processor and primary memory Tandem structure have been introduced, (Figure A-lb). Performance can be in- distribution channels; and even the including models by Auragen and Com- creased for these systems by having func- puter Consoles Inc. final user. tional multiprocessors serve disks and The price range of micros from $500 micro industry offers much The a terminals, including the migration of for a lap PC to nearly $500,000 for a fully wider range of computing products at software for file access. configured multimicroprocessor is much a lower cost-($500 to $500,000) than A completely symmetrical multipro- greater than that for any previous the mini ($20,000 to $500,000) or cessor can be made using more recent generation or computer class (see Figure mainframe ($250,000 to $5,000,000) buses such as the Multibus I1 or VME 1 in main article). Table A-1 illustrates industries can afford, and the micro bus (Figure A-lc), if a cache is used to the range of several Motorola 68000/ offers comparable performance. The reduce processor/memory traffic. The Unix-based computers that compete with results? A continued shakeout of all Aretk quad processor, which uses this the minicomputer.' types of products and companies and principle, is shown in Figure A-Id. Winners and losers in products, A variety of approaches are used to in- changing roles for all parts of the in- organization, and marketing may already crease system availability. Parallel com- be established.' However, many micro- dustry, including the users. puters (Figure A-le) use the Multibus for based products are still to be invented intercommunicating among distinct, outside the computer classes previously redundant computers (PC-Mp) and described. The box on the following page among redundant controllers for secon- contains questions about each structure dary memory (Ms) and terminals. Of in terms of competitiveness, long-term course, much software is required to pro- References stability, and substitution with other vide true high-availability computing structures. I. H. Hecht, 6'Computer Standards,,, with this structure. The structure is a In addition to these questions about Computer, vol. 17, N~.10, act. vastly scaled down version of the Tan- word processing, workstations, super- 19841 dem (Figure A- 1h). micros, and clusters of micros and high-

October 1984 availability computers is the most impor- tant question: that of standards, espe- cially Unix.

Uniw. For awhile, Unix appeared to be suitable only for a particular class of ex- perimental uses, but now it promises to be a constraint for the whole market. In- teractive computing with Unix is the product constraint future users are all hastening to demand, or at least specify. Just as the PC market has standardized on the IBM PC (8086, MS/DOS, PC bus, graphics interface, file formats, etc.), the market for systems larger than a PC appears to be standardizing on Unix. IBM has shown its flexibility in adopting industry standards especially when the time to market is crucial and the market demands it. If customers want a product, IBM will likely supply it. IBM has already announced Unix on the PC and will probably respond with Unix on its 4300 and mainframes. In a similar fashion, every minicom- puter and microcomputer supplier ap- pears to be offering Unix in a commodity-like fashion. While the com- bined market is large, the fundamental market has not been expanded, but merely made more accessible by every manufacturer. The result will be that more small manufacturers who have in- adequate marketing and manufacturing organizations will fail to compete with mainframe and mini suppliers. Unix has been an opiate that hundreds of companies have used as an excuse to form and assemble-quite trivially-a product from boards, Unix ports, or general-purpose software. Perhaps the entry cost for computer systems should be higher.

Office and word-processing systems. Historically, general-purpose computers have won in the marketplace over equiv- alent special-purpose machines. The IBM PC standard is the unique structure to watch as conventional word-proces- sing software becomes available and replaces simple editors. Terminals, in- cluding typewriters with built-in modems or computing telephones, can be con- nected to desktop and pedestal-sized, shared micros running Unix or to large systems for the casual users. Profes- sionals who already have large worksta- tions use them for text processing.

Workstations. Over 100 workstation vendors value themselves at up to $100

COMPUTER billion for a commodity-like product Table A-1. Selected 68000AJnix computer systems. with a limited market to engineers, scientists, and business analysts. All FIRST ENTRY MAXIMUM have enough organizational overhead to SYSTEM BUS STRUCTURE DELIVERY PRICE (K$) USERS start, but few have the critical mass or Apple Macin- ability to raise the next round of capital tosh Ext. serial PC to gain a significant market share except Corvus Uniplex Backplane Micro, LAN those well on the way-Apollo, Apple serve (with Macintosh), Convergent Technol- Altos 586-10 Multibus Shared micro ogy, and Sun-or those with unique high-performance products such as Sili- Wicat 150WS Backplane PC, shared con Graphics.' NCR Tower Workstation design consists of "assem- 1632 Multibus + Shared PcMp bus bling" the following: Plexus P/60 Multibus + Supermicro boards with microprocessors, disk, PcMp bus CRT, and communication con- SUN Worksta- trollers that use one of several stan- tion Multibus Workstation dard buses, such as Multibus, Qbus, ONYX C8002 None Shared micro or VME/Versabus; appropriate disks and CRTs; Aretk 1000 Single prop. Symmetric mP standard or custom enclosures; bus a licensed version of Unix available Synapse' Dual Symmetric high avail from myriad suppliers; and mP-:multi generic software, including word Stratus/32* Dual voting processing and spreadsheet. bus Fault tolerant Auragen 4000 Modified VME Multiprocessor Each start-up company believes its dual inter C Multicomputer, product and business plan will beat Tandem type Apollo, the first entrant into the high- performance, clustered workstation 'Operating system kernel LS not Unix-based market. In fall 1983, just after going tUNlX Review, June/July 1983. public, Apollo was valued at $1 billion $Degree of range for a multiprocessor. with annualized sales of less than $100 million and with fewer than 1000 employees. At the same time, Digital had Super-micro and clustered super- Because a somewhat different struc- a valuation of about $4 billion with sales micro systems. Basically this structure ture is involved in building high-avad- of $4 billion and a work force of over competes with old-line mini and main- ability computers, especially with respect 70,000. frame makers, both of which are begin- to software, there is a clear market. As ning to distribute supermicros (the Con- the overall reliability of computers in- A typical workstation start-up com- vergent Technology distribution model, creases, tne demand and pnce premium pany compares itself with Apollo on two for example). CT supplies hardware to for high-reliability or high-availability points: the start-up date (usually one to traditional manufacturers who use only computing is unclear. two years after Apollo when systems their distribution capability. Neither There is still interest in making a self- were easier to build) and the current group will let its base erode without diagnosable, self-repairing computer that month's annualized shipments. In this resistance, and both are ultimately never fails, however. While this feat is context, within two years, each of 100+ capable of backwardly integrating OEM possible for the CPU portion of a sys- companies will be valued at $1 billion hardware. tem, the peripherals and software do not dollars, giving a valuation of workstation permit the ultimate machine to be built companies of $10 to $100 billion. . .at for some time. least one order of magnitude greater than High-availability computer systems. The most important aspect of high- any optimistic projection of the market. High-availability computing, pioneered availability computers is that they can be This valuation doesn't include estab- by Tandem, may no longer be treated as designed for incremental upgrades using lished companies. The workstation is a a niche, but rather something a user both the multiprocessor and multicom- mainline product for large suppliers such should be able to trade off. Tandem's puter structures. This capability is why as AT&T (via new teletype computing product line is based on mini technology many computers are sold, regardless of terminals), DEC, HP, and IBM. Also, and as such now has about 20 companies their availability. With much lower the 32-bit personal computers circa targeting its base using microprocessors. priced machines, a broader range, and 1984-85, led by IBM using 256K chips DEC has introduced the Vax clusters in the introduction of fully distributed and the Intel 286, will provide the power the "Tandem-price" market, but VLSI computing in LAN clusters, the need for of emerging 68000-based Unix worksta- will reduce the cost. An IBM product is high-availability computers for in- tions at a lower price. long overdue. cremental expansion may decline.

October 1984 References

1. R. Baily, R. Scott, and K. Roberts, "Buyer's Guide To Hardware," Unix Review, Vol. I, No. I, June/July 1983, pp. 48-73. 2. S. T. McClellan, The Coming Com- purer Indusfry Shakeout, John Wiley & Sons, New York, 1984. 3. C. Machover and W. Myers, "lnterac- tive Computer Graphics," Computer, Vol. 17, No. 10, Oct. 1984.

C. Gordon Bell is chief technical officer for Encore Computer Corporation, where he is responsible for the overall product strategy. Before joining Encore Computer, he was vice president of engi- neering for Digital Equipment Corpora- tion, responsible for R&D activities in computer hardware, software, and sys- tems. He was also manager of computer design at DEC, responsible for the PDP- 4, -5, and -6 computers and served on the faculty of Carnegie-Mellon University from 1966 to 1972. Bell led the team that conceived the Vax architecture, established Digital Computing Architecture, and was one of the principal architects of C.mmp (16 processors) and Cm* (50 processors) at Carnegie-Mellon. He is widely published in computer architecture and computer design. Bell earned his BS and MS in electrical engineering at the Massachusetts In- stitute of Technology and holds several patents in computer and logical design. His address is Encore Computer Corp., 15 Walnut St., WeHesley Hills, MA 02181.