The clock rate of a CPU is normally determined by the frequency of an oscillator crystal. Typically a crystal oscillator produces a fixed sine wave—the frequency reference signal. Electronic circuitry translates that into a square wave at the same frequency for digital electronics applications (or, in using a CPU multiplier, some fixed multiple of the crystal reference frequency). The clock distribution network inside the CPU carries that clock signal to all the parts that need it. An AD Converter has a "clock" pin driven by a similar system to set the sampling rate. CPU manufacturers typically charge premium prices for CPUs that operate at higher clock rates, a practice called binning. For a given CPU, the clock rates are determined at the end of the manufacturing process through actual testing of each CPU. CPUs that are tested as complying with a given set of standards may be labeled with a higher clock rate, e.g., 1.50 GHz, while those that fail the standards of the higher clock rate yet pass the standards of a lesser clock rate may be labeled with the lesser clock rate, e.g., 1.3 GHz, and sold at a lower price. Chip manufacturers publish a "maximum clock rate" specification, and they test chips before selling them to make sure they meet that specification, even when executing the most complicated instructions with the data patterns that take the longest to settle (testing at the temperature and voltage that runs the lowest performance).
With any particular CPU, replacing the crystal with another crystal that oscillates half the frequency ("underclocking") will generally make the CPU run at half the performance and reduce waste heat produced by the CPU. Conversely, some people try to increase performance of a CPU by replacing the oscillator crystal with a higher frequency crystal ("overclocking"). However, the amount of overclocking is limited by the time for the CPU to settle after each pulse, and by the extra heat created.
After each clock pulse, the signal lines inside the CPU need time to settle to their new state. That is, every signal line must finish transitioning from 0 to 1, or from 1 to 0. If the next clock pulse comes before that, the results will be incorrect. In the process of transitioning, some energy is wasted as heat (mostly inside the driving transistors). When executing complicated instructions that cause many transitions, the higher the clock rate the more heat produced. Transistors may be damaged by excessive heat.
The first electromechanical general purpose computer, the Z3 operated at a frequency of about 5–10 Hz. The first electronic general purpose computer, the ENIAC, used a 100 kHz clock in its cycling unit. As each instruction took 20 cycles, it had an instruction rate of 5kHz.
The first commercial PC, the Altair 8800 (by MITS), used an Intel 8080 CPU with a clock rate of 2 MHz (2 million cycles/second). The original IBM PC (c. 1981) had a clock rate of 4.77 MHz (4,772,727 cycles/second). In 1992, both Hewlett-Packard and Digital Equipment Corporation broke the difficult 100 MHz limit with RISC techniques in the PA-7100 and AXP 21064 DEC Alpha respectively. In 1995, Intel's P5 Pentium chip ran at 100 MHz (100 million cycles/second). In March 6, 2000, AMD reached the 1GHz milestone a few months ahead of Intel. In 2002, an Intel Pentium 4 model was introduced as the first CPU with a clock rate of 3 GHz (three billion cycles/second corresponding to ~3.0×10−10seconds or 0.3 nanoseconds per cycle). Since then, the clock rate of production processors has increased much more slowly, with performance improvements coming from other design changes. A nano second is the time for light or an electric-signal to travel a distance of about 30cm, coming close to the distances a signal travels in the computer.
As of 2011[update], the Guinness Record for fastest CPU is by AMD with a Bulldozer based FX chip "overclocked" to 8.308 GHz; however, it has now been superseded by the next generation of AMD's FX chips "Piledriver" with a clock rate of 8.429 GHz.
Engineers continue to find new ways to design CPUs that settle a little more quickly or use slightly less energy per transition, pushing back those limits, producing new CPUs that can run at slightly higher clock rates. The ultimate limits to energy per transition are explored in reversible computing, although no reversible computers have yet been implemented.
Engineers continue to find new ways to design CPUs so that they complete more instructions per clock cycle (achieving a lower CPI count), although it may run at the same or a lower clock rate as older CPUs. This is achieved through architectural techniques such as instruction pipelining and out-of-order execution which attempts to exploit instruction level parallelism in the code.
The clock rate of a CPU is most useful for providing comparisons between CPUs in the same family. The clock rate is only one of several factors that can influence performance when comparing processors in different families. For example, an IBM PC with an Intel 80486 CPU running at 50 MHz will be about twice as fast (internally only) as one with the same CPU and memory running at 25 MHz, while the same will not be true for MIPS R4000 running at the same clock rate as the two are different processors that implement different architectures and microarchitectures. There are many other factors to consider when comparing the performance of CPUs, like the clock rate and width of the CPU's data bus, the latency of the memory, and the cache architecture.
The clock rate alone is generally considered to be an inaccurate measure of performance when comparing different CPUs families. Software benchmarks are more useful. Clock rates can sometimes be misleading since the amount of work different CPUs can do in one cycle varies. For example, superscalar processors can execute more than one instruction per cycle (on average), yet it is not uncommon for them to do "less" in a clock cycle. In addition, subscalar CPUs or use of parallelism can also affect the performance of the computer regardless of clock rate.