First of all, it will quickly become apparent to you that there’s little or no point in talking about fast performance here. Speed is not the yardstick to measure memory by, since any well-made PC100 SDRAM DIMM will provide performance nearly identical to any other such DIMM. In real-world terms, this is hardly anything to judge memory by. What you should expect from quality memory is stable system operation. Not that this is a yardstick you can use to separate the wheat from the chaff, either -most SDRAM modules will perform fairly equally at normal memory bus speeds (66MHz or 100MHz, depending on your system). That’s the whole point- these modules were designed to be stable at these frequencies. Stability problems will only arise if you’re thinking about overclocking -but more on that later.
 

ECC 

In the bad old days of SIMMs, error checking was at a comparatively primitive level and used an error-checking mechanism called parity. This used an extra, ninth parity bit for each 8 bits stored. Parity checking consisted of adding the values of the eight bits and storing either a 1 (if the result is even) or an 0 (if the result is odd) in the ninth bit. (This is called “even” parity, and the opposite system, “odd” parity, can also be used. This would simply mean that an odd sum would be stored as a 1 and an even sum as an 0.) For example, if the 8-bit packet was 10110001, the sum equals 4, which is even. If even parity is used, then the parity bit would have a value of 1. If any single bit changes its value, the parity check would note that the sum is now odd, not even, and recognise it as an error and halt the PC. This whole process does, admittedly, sound clunky and messy - and unreliable, in that if two bits changed value in the example above, the sum will remain even and the parity check would not recognize an error. But it wasn’t quite as messy as that, since even single bit errors occurred quite rarely and the probabilities were quite low.

Nowadays, the whole thing is more elegant. DIMMs use a feature called Error Correcting Code, or ECC, which gives you one byte of parity checking and automatically detects errors and is capable of even correcting single-bit errors without halting the PC. ECC memory is not slower than non-ECC memory and additional cycles occur only when errors are enountered. You need ECC DIMMs and a motherboard that supports this feature in order to use ECC. Of course, ECC is not an absolutely essential feature for most PCs, because data is often not irreplaceable and errors occur very seldom. For file servers, however, ECC is an enormously valuable option, because even small errors can be multiplied down the line and end up in large-scale and expensive operating failures.