Part of the nature of memory is that it inevitably fails. These failures are usually classified as two basic types: hard fails and soft errors. The best understood are hard fails, in which the chip is working and then, because of some flaw, physical damage, or other event, becomes damaged and experiences a permanent failure.

Fixing this type of failure normally requires replacing some part of the memory hardware, such as the chip, SIMM, or DIMM. Hard error rates are known as HERs. The other more insidious type of failure is the soft error, which is a nonpermanent failure that might never recur or could occur only at infrequent intervals. (Soft fails are effectively "fixed" by powering the system off and back on.) Soft error rates are known as SERs.

About 20 years ago, Intel made a discovery about soft errors that shook the memory industry. It found that alpha particles were causing an unacceptably high rate of soft errors or single event upsets (SEUs, as they are sometimes called) in the 16KB DRAMs that were available at the time.

Because alpha particles are low-energy particles that can be stopped by something as thin and light as a sheet of paper, it became clear that for alpha particles to cause a DRAM soft error, they would have to be coming from within the semiconductor material.

Testing showed trace elements of thorium and uranium in the plastic and ceramic chip packaging materials used at the time. This discovery forced all the memory manufacturers to evaluate their manufacturing processes to produce materials free from contamination.

Today, memory manufacturers have all but totally eliminated the alpha-particle source of soft errors. Many people believed that was justification for the industry trend to drop parity checking. The argument is that, for example, a 16MB memory subsystem built with 4MB technology would experience a soft error caused by alpha particles only about once every 16 years!

The real problem with this thinking is that it is seriously flawed, and many system manufacturers and vendors were coddled into removing parity and other memory fault-tolerant techniques from their systems even though soft errors continue to be an ongoing problem.

More recent discoveries prove that alpha particles are now only a small fraction of the cause of DRAM soft errors. As it turns out, the biggest cause of soft errors today are cosmic rays. IBM researchers began investigating the potential of terrestrial cosmic rays in causing soft errors similar to alpha particles.

The difference is that cosmic rays are very high-energy particles and can't be stopped by sheets of paper or other more powerful types of shielding. The leader in this line of investigation was Dr. J.F. Ziegler of the IBM Watson Research Center in Yorktown Heights, New York.

He has produced landmark research into understanding cosmic rays and their influence on soft errors in memory. One example of the magnitude of the cosmic ray soft-error phenomenon demonstrated that with a certain sample of non-IBM DRAMs, the SER at sea level was measured at 5950 FIT (failures in time, which is measured at 1 billion hours) per chip.

This was measured under real-life conditions with the benefit of millions of device hours of testing. In an average system, this would result in a soft error occurring every six months or less. In power-user or server systems with a larger amount of memory, it could mean one or more errors per month!

When the exact same test setup and DRAMs were moved to an underground vault shielded by more than 50 feet of rock, thus eliminating all cosmic rays, absolutely no soft errors were recorded. This not only demonstrates how troublesome cosmic rays can be, but it also proves that the packaging contamination and alpha-particle problem has indeed been solved.

Cosmic-ray-induced errors are even more of a problem in SRAMs than DRAMS because the amount of charge required to flip a bit in an SRAM cell is less than is required to flip a DRAM cell capacitor. Cosmic rays are also more of a problem for higher-density memory.

As chip density increases, it becomes easier for a stray particle to flip a bit. It has been predicted by some that the soft error rate of a 64MB DRAM will be double that of a 16MB chip, and a 256MB DRAM will have a rate four times higher.

Unfortunately, the PC industry has largely failed to recognize this cause of memory errors. Electrostatic discharge, power surges, or unstable software can much more easily explain away the random and intermittent nature of a soft error, especially right after a new release of an operating system or major application.

Studies have shown that the soft error rate for ECC systems is on the order of 30 times greater than the hard error rate. This is not surprising to those familiar with the full effects of cosmic-ray-generated soft errors. The number of errors experienced varies with the density and amount of memory present.

Studies show that soft errors can occur from once a month or less to several times a week or more! Although cosmic rays and other radiation events are the biggest cause of soft errors, soft errors can also be caused by the following:

• Power glitches or noise on the line. This can be caused by a defective power supply in the system or by defective power at the outlet.

• Incorrect type or speed rating. The memory must be the correct type for the chipset and match the system access speed.

• RF (radio frequency) interference. Caused by radio transmitters in close proximity to the system, which can generate electrical signals in system wiring and circuits. Keep in mind that the increased use of wireless networks, keyboards, and mice can lead to a greater risk of RF interference.

• Static discharges. Causes momentary power spikes, which alter data.

• Timing glitches. Data doesn't arrive at the proper place at the proper time, causing errors. Often caused by improper settings in the BIOS Setup, by memory that is rated slower than the system requires, or by overclocked processors and other system components.

Most of these problems don't cause chips to permanently fail (although bad power or static can damage chips permanently), but they can cause momentary problems with data. How can you deal with these errors?

Just ignoring them is certainly not the best way to deal with them, but unfortunately that is what many system manufacturers and vendors are doing today. The best way to deal with this problem is to increase the system's fault tolerance. This means implementing ways of detecting and possibly correcting errors in PC systems.

Three basic levels and techniques are used for fault tolerance in modern PCs:

• Nonparity

• Parity

• ECC

Nonparity systems have no fault tolerance at all. The only reason they are used is because they have the lowest inherent cost. No additional memory is necessary, as is the case with parity or ECC techniques. Because a parity-type data byte has 9 bits versus 8 for nonparity, memory cost is approximately 12.5% higher.

Also, the nonparity memory controller is simplified because it does not need the logic gates to calculate parity or ECC check bits. Portable systems that place a premium on minimizing power might benefit from the reduction in memory power resulting from fewer DRAM chips.

Finally, the memory system data bus is narrower, which reduces the amount of data buffers. The statistical probability of memory failures in a modern office desktop computer is now estimated at about one error every few months. Errors will be more or less frequent depending on how much memory you have.

This error rate might be tolerable for low-end systems that are not used for mission-critical applications. In this case, the extreme market sensitivity to price probably can't justify the extra cost of parity or ECC memory, and such errors then must be tolerated.

At any rate, having no fault tolerance in a system is simply gambling that memory errors are unlikely. You further gamble that if they do occur, memory errors will result in an inherent cost less than the additional hardware necessary for error detection. However, the risk is that these memory errors can lead to serious problems.

A memory error in a calculation could cause the wrong value to go into a bank check. In a server, a memory error could force a system to hang and bring down all LAN-resident client systems with subsequent loss of productivity.

Finally, with a nonparity or non-ECC memory system, tracing the problem is difficult, which is not the case with parity or ECC. These techniques at least isolate a memory source as the culprit, thus reducing both the time and cost of resolving the problem.