Memory Speed

Memory speed is one of the most important limitations on the performance of PCs. Affordable memory chips simply cannot keep up with the processing speed of today's Pentium processors.

In fact, speed deficiencies are nothing new, having first appeared when IBM introduced its original AT computers with their 80286 microprocessors. Ordinary memory chips could not keep pace with the speed of such a fast (by the standards of 1984, remember) microprocessor. The 80286 could request bytes in such short order that memory was unable to respond. Consequently, IBM added what has become the bane of every memory system in PCs to this day, wait states.

A wait state is exactly what it sounds like; the microprocessor suspends whatever it's doing for one or more clock cycles to give the memory circuits a chance to catch up. The number of wait states required in a system depends on the speed of the microprocessor in relation to the speed of memory.

Microprocessor speeds are usually expressed as a frequency in megahertz-millions of cycles per second-while memory chips are rated by time in nanoseconds-billionths of a second. The two measurements are reciprocal. At a speed of one megahertz, one clock cycle is 1000 nanoseconds long; eight megahertz equals 125 nanoseconds; sixteen megahertz, 62.5 nanoseconds; twenty megahertz, 50 nanoseconds; twenty-five megahertz, 40 nanoseconds; thirty-three megahertz, 33 nanoseconds; and so on.

Dynamic memory chips are speed-rated; usually with a number emblazoned on the chip following its model designation. This number reflects the access time of the chip given in nanoseconds with the rightmost zero left off to make the expression a little more compact. A chip that has a -12 labeled on it, therefore, has an access time of 120 nanoseconds.

If this were the number of merit for chip speed, most of today's computers would have no problem. At 25 megahertz, for instance, one clock cycle is 40 nanoseconds and the microprocessors require at least two cycles between memory operations, a total of 80 nanoseconds. Chips rated at 70 nanoseconds are readily available and relatively inexpensive. In general, you'll do no harm installing quicker chips than a computer calls for, for instance putting 70 nanosecond chips into a system that calls for 80 nanosecond parts. The only detriment is that faster chips will likely cost you more-you'll be paying for speed that you don't need. Slower chips may not work or, more likely, work sporadically, leaving you vulnerable to parity check errors at unexpected times.

The access time is not the only-or the most important-figure to describe a memory chip, however. More relevant is the cycle time, which does measure how quickly two back-to back-accesses can be made to the chip. The cycle time is generally about two to three times the access time of the chip. Even an 70-nanosecond DRAM chip, therefore, is not capable of reliably serving a 25 MHz PC.

Static RAM chips have no need to be refreshed. Not only do their cycle times equal their access times, but they can operate faster. Static chips are readily available with ratings of 25 or 35 nanoseconds while the fastest common DRAM chips are rated at 60 or 70 nanoseconds. Unfortunately, because static chips are much more expensive than DRAM, they are rarely used for the primary storage of PCs.

To cope with the speed limitations of affordable DRAM memory chips, PC makers use a number of designs for their memory system. The two most straightforward of these are simply to use the fastest possible chips-but even today's quickest DRAM chips lag far behind a 50 or 66 MHz microprocessor-or to impose as many wait states as necessary (with not-so-quick results). A single wait state extends a normal memory cycle from two to three clock ticks-that's a big performance hit. With one wait state, a PC operates at only two-thirds its potential speed. Two wait states cut performance in half.