Microprocessor Families

Developments in computer systems don't always follow logically from the underlying technology. In many cases, change is accelerated or slowed according to perceptions of the business environment. A brief description of the major microprocessor manufacturers will illustrate this relationship.

Intel Corporation

Intel, established in 1968, created the world's first microprocessor, the 4004, in 1971. Intel's founder, Robert Noyce, was a co-inventor of the integrated circuit, and a one-time protégé of William Shockley, one of the inventors of the transistor. Gordon Moore and Andrew Grove were other key principals of Intel from the earliest days.

Intel's first microprocessor was made under contract for a Japanese company that wanted it for a calculator. The chip featured 2,300 transistors. Although Intel continued to be successful in the microprocessor market, it was not the technology leader in 1980 when IBM went looking for a CPU for its first personal computer. Motorola's just introduced 68000 chip was superior in nearly every respect to Intel's existing flagship, the 8086. But IBM was not looking for the leading edge, as evidenced by the fact that it chose not the 16-bit inside/16-bit data bus 8086, but the less capable 16-bit inside/8-bit data bus 8088. IBM was in a hurry to get its machine to the market, and components for 8-bit buses were more widely available than 16-bit.

Just as important as technology was the business side. Intel, in 1981, was already the stable, successful kind of company that IBM favored; its sales had reached $300 million in 1978. Even so, IBM followed a corporate precedent in insisting that Intel license the right to manufacture its chip, as well as succeeding generations, to IBM and to other fabricators. This insistence on "second sourcing" of key components was standard procedure for many computer manufacturers.

Table 4.1 summarizes the generations of Intel CPUs.

Table 4.1. Intel's desktop CPUs. It isn't possible to accurately compare the abilities of these chips across all generations, but on standard benchmarks, the 650 MHz Pentium III is about seven times faster than the 200 MHz Pentium.

Intel CPUs
Generation	Year	Name	Transistors	Initial speed	Special features
I	1979	8088	29,000	4.77 MHz	* 8086 version had 16-bit data bus
II	1982	80286	134,000	6 MHz	*Added 24-bit address bus
III	1985	80386	275,000	20 MHz	* First true 32-bit chip
					* Intel dropped "second sourcing"
IV	1989	80486	1.2 million	25 MHz	* Clock doubling
					* On-chip cache
					* On-chip floating point unit
V	1993	Pentium	3.1 million	60 MHz	* Superscalar; two instructions per clock cycle
VI	1995	Pentium Pro	5.5 million	150 MHz	* Longer pipeline (14 stages vs. 5)
					* RISC techniques employed at heart of system
					* L2 cache in high-speed cartridge
VI	1997	Pentium II	7.5 million	233 MHz	* Variation of the Pentium Pro
					* Adds Integer SIMD abilities
					* Lots more L1 cache
VII	1999	Pentium III	9.5 million	450 MHz	* Adds floating point SIMD abilities
					* Reaches 733 MHz same year
VIIa?	2000	"Willamette"	?	1 GHz	* New flagship 32-bit CPU; note only about a year behind previous redesign
					* Significant improvement in FPU capabilities
VIII	2000	Itanium	?	800 MHz	* True 64-bit processor
					* VLIW-type technology

It'll Cost Ya The fastest chips in a family are disproportionately expensive. When AMD announced its Athlon (K7) line in August of 1999, this is how their chips were priced. Intel's marketing strategy is similar. AMD K7 Prices Speed Unit price 650 MHz $849 600 MHz $615 550 MHz $449 500 MHz $249 Source: EE Times on the Web.

One striking thing about Intel's history is the accelerating rate of change. With the debut of the Itanium, three distinct generations will have appeared in only about five years. Clock speeds increased from just over 200 MHz to over 700 MHz in less than three years. Given this progress, you would have to think that Intel at the turn of the century was more dominant than ever. In fact, the opposite is true. Intel's market share, which stayed above 90 percent for most of the 1980s and 1990s, dipped to around 80 percent in the last years of the 1990s. The reason for this change was the appearance of the sub-$1,000 PC.

A number of Intel's competitors, led by AMD, began to offer very low priced CPUs for the new class of sub-$1,000 PCs that appeared around 1998. Since Intel did not have a competitive chip, it lost market share as the new category of machine soared in popularity. The shift was a little deceptive however; low price also means low profit, and Intel continued to dominate the high end. To understand the difference in profit, see the sidebar.

So, there's been bad news and good news for Intel. Unfortunately, there's more bad news. An increasingly serious problem for the mainstream microprocessor vendors is what appears to be increasing resistance to high-end pricing. Computer buyers, especially corporate ones, have come to realize that faster processors secure only a very small improvement in the performance of real world applications software. Beginning in 1998, when mass market CPU speeds reached 300 MHz, improvements in software performance began to diminish and vendors began to notice an increasing indifference to faster processors. People simply didn't want to pay the large price premium, for example, the 340 percent shown in the sidebar, when going from 500 to 650 MHz. Sales overall have increased, but the market growth has now shifted to the low end.

Intel currently has three different microprocessor lines, with one more soon on the way.

The Celeron

The low-end Celeron, rushed to market in 1998 to fight inexpensive rivals from AMD and Cyrix, was at first essentially a budget version of the Pentium II. Intel has since moved to give it the same SIMD instructions as the Pentium III.

The Pentium III

Distinguished initially by the addition of a special floating point pipeline, the III is now the mainstream desktop chip. It differs from the Celeron largely in the amount of high speed cache.

The Xeon

The Xeon, which is aimed at servers and workstations, appeared in the second half of 1998. It has a Pentium III core and comes with logic that allows it to be used easily in systems that have as many as eight CPUs. Xeons also have more of the highest speed cache.

The Itanium

Scheduled to appear in 2000, the Itanium will be Intel's first 64-bit chip. Like the Xeon, which it will eventually replace, the Itanium is designed for servers and workstations.

Intel Clones

Intel CPUs were manufactured from the beginning by a number of licensees because of the need for second sourcing. Beginning with the 386, Intel decided that it would no longer license its designs. This didn't affect IBM, whose original agreement gave it the option to continue to manufacture Intel designs if it so chose (IBM exercised this right only sporadically). Rival AMD thought it also had a continuing right to fabricate everything in the Intel's x86 family and sued to enforce it. The results of the lawsuit were mixed, but AMD has grown to be Intel's principal challenger and we'll turn to it next.

AMD and NexGen

AMD (Advanced Micro Devices) was the most aggressive of the cloners during the 386 and 486 years. Most observers agree that AMD's success with these chips was an important factor in causing Intel to advance the development and marketing of the Pentium. AMD's cloning took two paths; the first was a direct copy of the 386, including on-chip software known as microcode (see Chapter 5), which Intel challenged in court. While the mutual lawsuits wound their way through the legal system, AMD buttressed its clone with a second approach. It created a version of the 486 that was not an exact copy of the Intel original. Instead, the microcode was written by engineers who had no knowledge of the real thing. This "reverse engineering" approach had been pioneered earlier in the development of BIOS chips (also discussed in Chapter 5) and had withstood legal challenges. The reverse engineered 486 chips were very successful technically and modestly competitive in the market.

Where Your Chip Dollars Go The legal struggle between AMD and Intel, which lasted from 1987 to 1995, is estimated to have cost some $200 million for lawyers and their expenses. The results were inconclusive, in large part because changes in the technology had made many of the issues moot.

AMD became more competitive with the sixth (Pentium Pro/Pentium II) generation. The advantage for AMD was that Intel had itself engaged in a form of reverse engineering in order to graft their old CISC system into a new RISC-based one. AMD could now simply find its own ways to do the same thing. With the help of intellectual property gained through the acquisition of fellow cloner NexGen, AMD has had a string of very competitive chips. Unfortunately, the company has chronically failed to manufacture its newest chips in volume large enough to threaten Intel's leadership. Taken with the fact that, as a cloner, AMD has to charge less than Intel for a comparable chip, this has not been good for the company's profits. AMD has taken a chunk out of Intel's market share, but only in the low-profit, low end of the market. AMD's seventh generation chip, the Athlon (K7) introduced in 1999, has looked very competitive against the comparable high-end Pentium III, but to make a dent in the high end, AMD must resolve its traditional problems in getting sufficient volume to the market. Efficient, low-cost manufacturing has been one of Intel's strengths.

Cyrix

A Texas-based company, Cyrix Semiconductor was the first to make a clone of an Intel CPU without using Intel microcode. Cyrix was seduced into the market by Intel's huge profit margins and felt that it could compete by going fabless—IBM and others provided its foundries. Still, Cyrix always struggled to make a profit. It is now owned by the Taiwanese chipset and motherboard manufacturer Via Technologies. A similar fate has befallen other cloners, such as Rise and IDT.

RISC CPUs

The CPUs that follow are fundamentally different from those discussed so far in that they support different software, both operating systems and applications, than the Intel family. It is possible to compare their performance to the Intel and Intel clone CPUs on abstract benchmarks, but these don't really give a fair indication of their relative power. As we'll discuss in Part Two, the relationship of hardware to software is a complex one. For example, CPUs rely on software techniques (optimizing compilers) to maximize the performance of programs written for them. It is extremely difficult to separate the performance of these compilers from the performance of the CPU itself. Further complicating the question of which is most powerful is the fact that almost all of the new RISC CPUs are 64-bit processors. Since there is very little standardized software around that is written for such an incredibly wide processing path, real comparisons of these CPUs to each other, not to mention to the 16- and 32-bit world, will remain difficult for quite some time.

IBM and Motorola: PowerPC

The PowerPC architecture is the first in this list that is not compatible with the Intel family; its instruction set, external connections, etc. are completely different. The PowerPC is a true RISC approach, building on IBM's early experience, including especially the RS6000 workstation.

The PowerPC alliance came about when a number of factors converged. IBM was looking for a mass market for its CPU technology and Apple was looking for an alternative to the Motorola 68xxx family of CPUs that had powered its Macs since they first appeared in 1984. The Motorola 68000, used in that first Mac, was the best CPU of its time. It operated with 32 bits internally and 16 externally; its "flat" memory addressing was much superior to the awkward segmented scheme used in Intel's chips. Apple did well with successor chips, from the 68020 (1985), the 68030 (1987), to the 68040 (1989). But, even before the 68050 could come to market, it was clear to Apple engineers that Motorola's 68K architecture was running out of steam. Apple had to do more than keep up with the "WinTel" (Windows/Intel) system; it had to offer something better if it was to overcome its competitors' tremendous momentum. RISC seemed the only answer. So IBM, Motorola, and Apple talked and Apple agreed to adopt a chip based on IBM's "Power" architecture—the PowerPC. Although derived from an IBM design, Motorola's engineers have made important contributions to the PowerPC series and Apple's participation was significant as well.

Emulation You may have wondered if it's possible to have software written for one chip work on another type that isn't a clone—that has a different instruction set. Indeed this can be done; it's called emulation. There are two approaches: software and hardware. On the soft side, emulation means an additional layer of software that, for example, traps other programs' requests to the Intel instruction set and converts them to that of the PowerPC. The penalty here, of course, is that the computer is running another program, one that requires a share of the available CPU cycles, cache, and memory. Software emulation is typically slow, but it is used, for example, by the PowerPC in emulating the Motorola 68K CPUs for older Macintosh software. Emulation in hardware does the same thing, but uses extra circuits on the chip instead of added software. This is obviously much faster. We've seen that Intel's sixth generation and later chips actually emulate older Intel chips. But, aside from this limited example, there hasn't been much effort to make CPUs that run two completely different instruction sets. The reason is that the penalty in transistors is too high. As advancing technology begins to permit a billion or so transistors on a single chip, that could change. In the meantime, however, even Transmeta's state-of-the-art Crusoe processor does its emulation in software.

The PowerPC has been a technological success but a commercial failure on the desktop. Despite the performance boost supplied by the chip when it appeared in the early 1990s, Apple's market share declined (it has since rebounded). IBM and Motorola were unable to balance this loss of volume with other system adoptions. The current version of the PowerPC, the G4, is a very competitive CPU, but overall sales have been so low that IBM has dropped out of the alliance. Motorola is now responsible for all design and manufacturing. Variations of the PowerPC sold by both companies are quite strong in the embedded processor market, though.

Silicon Graphics: MIPS

The MIPS chip family derives from research done at Berkeley and Stanford by faculty who did the usual Bay Area thing—they got some venture capital, left academe, and founded a start-up. The need for cash was minimized by going fabless. Key licensees were NEC, Sony, and Siemens; NEC remains a major manufacturing partner. Silicon Graphics (SGI) bought the company in 1992 when it was struggling; SGI's move was as much self-protection as anything. SGI had settled on MIPS chips for its workstations and didn't want to be forced to change. But, with a small market share and little income to invest in R&D, MIPS floundered and SGI sold it off again after a few years. The MIPS architecture will still be around for a long while, however. Versions of the various MIPS "cores" are widely licensed and used in embedded processors and specialized CPUs.

Sun: SPARC

Sun, a maker of workstations that originally used the Motorola 68K architecture, designed its own RISC chip to move up in the incessant contest for more speed that characterizes the workstation market. Sun's SPARC (scalable processor architecture) specifications were licensed to other fabricators and were realized in silicon in a variety of designs. The key participants were technological powerhouse Texas Instruments, as well as Japan's Fujitsu. For a time in the 1980s, Sun appeared to be on the way to dominating the workstation market. Traditional rivals Compaq/DEC, HP, and IBM were off balance and seemingly in retreat. But Sun's alliance lagged behind the others in developing more powerful versions of the CPU and by the early 1990s, Sun was at the bottom of the performance heap. The company as a whole sagged, but it didn't fold because of a very strong base of loyal customers. Sun still has about a third of the workstation market and its current UltraSPARC CPU is strong in servers and routers. Sun is also taking the lead in designing CPUs to run programs written for the Java environment (see Chapter 9 for a discussion of Java). This could make Sun a player in the huge market for embedded processors.

Sun over Wall Street Sun, among the brashest of Silicon Valley companies, has perhaps its strongest following on Wall Street. In the late 1980s, Sun's workstations effected a sort of revolution in the stock market as they allowed technology-oriented traders to program one workstation to link multiple information systems. New kinds of analysis and new kinds of trading then followed. The traders also regained "desk real estate" as they got rid of the multiple terminals that they once had to contend with. When Sun's relative bang for the buck slipped in the early 1990s, these customers remained loyal and continue to be so today.

Compaq/DEC: Alpha

Digital Equipment Corporation (DEC, now owned by Compaq) designed and fabricated most of its own CPUs from the days of its first minicomputer, the PDP, in 1959. It took DEC's founder and CEO, Kenneth Olsen, a while to realize that his CISC architecture (the VAX) was going to come to a dead end (like a lot of people who begin careers with remarkable vision, Olsen failed to adapt to change in his later years at the helm). DEC's first move was to invest heavily in a RISC architecture that it then jettisoned when it was almost ready; the experiment with MIPS was the alternative. But DEC had some talented designers and they eventually got management to sign on to a super speed RISC project—the Alpha.

The Alpha series of CPUs, which now sit at the heart of all of Compaq's workstations, servers, and small mainframes, is exceptionally powerful. More than any other company's offerings, the Alpha emphasizes clock speed. Alpha CPUs are 64-bit throughout, employ multiple deep pipelines, and very large on-chip caches, but do not support out of order execution. Compaq prefers raw speed to special tricks. One disadvantage of all that speed is that the chips run much hotter than the competition. If Compaq made an Alpha portable, it would be very uncomfortable to its user's lap for the very short time before its battery died.

Compaq has embraced Microsoft's Windows NT/2000 as a way of opening its systems to the vast array of Windows-based software. This move has been reinforced by making the Alpha CPU compliant with industry standard parts—the PCI bus, the ATX motherboard, and so forth, and by the fact that Microsoft has translated some of its most successful applications (Word, Excel, other elements of Microsoft Office) to run directly (without emulation) on Alpha machines. Together, these moves are designed to encourage the production of Alpha machines from vendors other than Compaq. So far, there have been few takers, though Korean semiconductor giant Samsung has licensed the technology and promises to advance its capabilities. The combination of Compaq's market scale, the former DEC's R&D, and the fabricating skills of Samsung could make a formidable combination.

HP: PA-RISC

Hewlett-Packard (HP) is a major player in the workstation and server markets with its PA (Precision Architecture) RISC family. The current leader, the PA-8000, is a true 64-bit CPU that has the usual state-of-the-art elements: it is superscalar, superpipelined, and as with all workstation-oriented chips, emphasizes floating point and advanced graphics performance. The race for leadership in workstation performance is a perpetual seesaw battle, with the newest entrant usually seizing the banner and holding it for a few months until the next new version of a competitive CPU is announced. The PA-8000 has had its share of honors, especially in floating point performance, but HP's alliance with Intel to design the Itanium suggests that HP considers the PA-RISC architecture to have reached the end of its natural life.

Can Anyone Compete with Intel?

Intel has some powerful advantages in manufacturing and marketing microprocessors. First, it has the software "legacy." To quote Carl Sagan, there are "billions and billions" of dollars worth of software programs out there that are designed to operate only with Intel's x86 architecture. Even if Intel's CPUs slipped behind in performance relative to the competition, it wouldn't be enough to cause all those software developers to go to the tremendous expense of rewriting their programs to run on a different CPU family (one of the problems they would have is which different CPU). So, Intel would have to not merely slip, but really fall behind, to make people reconsider its CPUs for either software development or purchase.

Intel's second advantage in the market is its famed skill at manufacturing. Unlike many big companies that get sloppy when they dominate a market, Intel has given fanatical, unrelenting attention to production efficiency and quality. To illustrate this, consider Intel's "Copy Exactly" process. Instead of first using a development facility to get processes correct and subsequently a production one to adapt these to volume, Intel fine tunes for both dimensions at the same time in the same fab. Once things are perfected, all of Intel's many fabs around the world copy the process as exactly as possible. This not only speeds development time, it also means that problem solving can be replicated throughout the system almost instantly. While Intel has been challenged at the low end of the market in recent years, and lost share as a result, its problem has not been the ability to compete at a profit, but rather just how much of its existing profits it wanted to sacrifice. Intel's profits are in fact extremely high. Before the recent fight with AMD began, its margins were in the range of 40 percent—amazing for a capital intensive business. And that leads us to Intel's third advantage, cash on hand—billions and billions again. If ever Intel is beaten, it won't be because it lacks the resources to invest.

So what are Intel's weaknesses? One is technology. Its CPU's have typically just equaled the most ambitious cloners, and have rarely been performance leaders vs. the RISC world. This is deceptive, however. Just as the legacy software is an advantage for Intel, it is also a problem; the company can't easily follow radical new design strategies because it always has to worry about "backwards compatibility." So far, it has done a great job of both staying compatible and keeping its best CPUs near the top in power. It has also consistently beaten everyone in price/performance ratio.

Intel's other weakness is the breadth and direction of the microprocessor market as a whole. You would expect a company with more than 80 percent market share to be engaging in regular discussions with flinty-eyed lawyers from the Justice Department. In fact, Intel has had very little scrutiny, largely because it is a manufacturer of microprocessors, not just CPUs for desktop computers. Viewed from this broader perspective, covering devices from mainframes to cell phones, Intel's share is strong but not dominant.

Thinking of the amazing array of devices that currently employ CPUs leads us to the next topic.