There are three ways to make a faster microprocessor: increase the clock speed; design a better microarchitecture; or increase the transistor density. All these methods are interrelated, and each carries its own set of trade-offs.
Boosting the clock speed allows a CPU to process instructions more quickly, but it also requires more wattage, dissipates more heat, and increases the cost of other system components. Designing a better microarchitecture improves throughput, but it can take years of R&D. Also, it usually requires more logic, which increases the transistor count, enlarges the chip's die size, and hikes the manufacturing cost. Increasing the transistor density enables higher clock speeds and better microarchitectures, but it also dissipates more heat, unless the voltage is decreased to compensate. Lower voltages require syste
m vendors to redesign their motherboards and chip vendors to redesign their peripheral components.
Microprocessor evolution typically follows this course: First, a better microarchitecture introduces a new generation of that CPU family. Later versions within that generation boost the clock speed and increase the transistor density (if not the actual transistor count) as smaller process technologies become available. Finally, when higher clock speeds and greater densities yield diminishing levels of return, the next-generation microarchitecture debuts, and the whole cycle starts over again.
Intel's 486 is a classic example. When introduced in 1989, the 486 offered a better scalar architecture than the 386, ran at 25 MHz, and was fabricated on a 1.0-micron process. Later, Intel shifted manufacturing to a denser 0.8-micron process and boosted clock speeds to 66 MHz. This year, Intel announced the 100-MHz IntelDX4 and moved production to a 0.6-micron line. Meanwhile, Intel also initiated another pro
duct cycle by introducing the next-generation Pentium.
This is a high-stakes game in which few companies are equipped to compete. A state-of-the-art, submicron fabrication plant (known as a fab) costs a billion dollars and takes years to build. Because a high-end microprocessor also takes years to develop, engineers have to target a fab process that isn't yet available. If everything does not come together at the right moment, the result can be financial disaster.
Some chip vendors dodge this problem by subcontracting all their fab work to outside suppliers. Cyrix and NexGen are "fabless" design houses that rely entirely on other semiconductor companies to manufacture chips. Cyrix now subcontracts its fab work to SGS-Thomson and IBM Microelectronics. The IBMdeal, announced in April, should assure Cyrix's customers that supply will meet demand. In addition, it gives IBM the right to produce an equal number of Cyrix-designed chips (including the M1 and its successors) for use in its own systems or
for sale to other system vendors.
Intel, AMD, and IBM Microelectronics have their own fabs, but they often fall short of demand. Intel and AMD are building new fabs, and AMD is also farming out work to DEC. Companies with excess capacity frequently make deals like this to recoup the huge costs of their fabs.
The capital investments for fabs are becoming so great that they almost dictate a semiconductor company's business strategy. In 1992, Intel ceded the still-healthy 386 market to AMD to produce more 486 chips, which command a higher price and are more profitable.
These business strategies should be kept in mind when comparing the price/performance ratios of competing microprocessors. For example, although RISC chips like the PowerPC 601 enjoy an approximate 2-to-1 price/performance advantage over the Pentium, it's not clear that it is entirely due to a corresponding difference in manufacturing costs. Both chips are now made on similar processes with similar transistor densities. Intel
may simply be charging higher prices to amortize its capital investments more quickly. All semiconductor companies must play by the same basic rules of physics and finance.
A Billion-Dollar Ball Game
