ECC SDRAM Primer

What is ECC SDRAM? ECC (error correction code) SDRAM is memory that is able to detect and correct some SDRAM errors without user intervention. ECC SDRAM replaced parity memory which could only detect, but not correct, SDRAM errors.

What causes SDRAM errors? Per Dell, "Memory errors are characterized as hard or soft. Hard errors are caused by defects in the silicon or metalization of the SDRAM package, and are usually permanent once they manifest. Soft errors are caused by charged particles or radiation, and are transient. In the past, soft errors were primarily caused by alpha particles, but that failure mode has been mostly eliminated today by strict quality control of the packaging material by SDRAM vendors. Currently the primary source of soft errors in SDRAM is electrical disturbance caused by cosmic rays, which are very high-energy subatomic particles originating in outer space."

A typical PC 133, 9 chip, 16 x 64, 128 MB ECC SDRAM.

What happens when a SDRAM crash occurs? When main memory crashes, all data in memory is lost. The larger the amount of main memory on the computer, the greater the possibility of nonrecoverable data loss.

What kind of errors can ECC SDRAM correct? Most ECC SDRAM can correct single bit errors, and detect, but not correct larger errors. Thus, errors greater in size than 1 bit will still crash the computer.

Chipkill was invented to augment ECC DRAM. Large server manufacturers have implemented additional error correcting hardware capabilities with a technology known as Chipkill.
Per Dell, "Chipkill correct is the ability of the memory system to withstand a multibit failure within a SDRAM device, including a failure that causes incorrect data on all data bits of the device. These methods rely on the chip set and hardware architecture of the system and cannot be achieved through software upgrades."

So what is the possibility of data loss? The data shown below illustrates the results of an IBM analysis comparing server outages due to memory failures of parity, ECC and Chipkill-equipped servers.

In summary, the following outage rates were identified:
A 32MB parity memory-equipped server received 7 outages per 100 servers over 3 years.
The 1GB ECC memory-equipped server received 9 outages per 100 servers over 3 years.
The 4GB Chipkill-equipped server received 6 outages per 10,000 servers over 3 years.
It can be seen that the Chipkill equipped sever had a failure rate of a magnitude of over 10 times lower than regular ECC SDRAM. Also, remember that the more system memory a computer has, the more likely it will crash due to a memory error.

What about speed? I could find no conclusive evidence that ECC SDRAM performed any slower than non-ecc SDRAM. Both Dell and IBM stated in their referenced articles there was no speed penalty to use a Chipkill enhanced server instead of an ECC memory equipped server without Chipkill.

So who should buy ECC SDRAM? First, the average user should be frequently saving data to their hard drive, so the likelihood of catastrophic memory failure should be small and therefore ECC memory would be overkill.
Second, if you are thinking of running a server, you definitely want to have a working RAID disk array, as your hard drives are much more likely to fail then your memory.
Third, if you want to run a server, there is no reason not to have ECC memory if your motherboard supports it. Currently ECC SDRAM only costs a little bit more than regular SDRAM.

Referenced Articles

IBM Chipkill Memory - IBM, February 1999

Chipkill Correct Memory Architecture - Dell, August 2000