Some new data that we have never had before.
We had some data from bulk testing of DDR5 RAM about the number of errors found per test and also our own data about how long each test takes to run.
The numbers are big enough to be statistically significant (>10,000 test runs).
The most effective test in terms of finding errors is "Moving inversions, 32bit pattern". But it is a relatively slow test.
The most effective test in terms of finding errors in the shortest time is the "Address test, own address".
Least effective test was Test 0 - Address test, walking ones. But we think this can be improved in a new release.
Note that no data was available for Test 6, block move, or 13 in this sample. But anecdotally 6 it is reasonably effective.
2nd note: Errors per second was total errors over all runs, while time was just a typical average single run test time. So the Errors / sec numbers look a lot bigger than you might expect, but are easier to read. (actual errors found per second on a single machine are more around, 0.0005 errors found per second of testing). This doesn't change the relative ranking however.
3rd note: These results are from a volume production environment. Not from retail sticks you would buy in a store. Failure rates are expected to be significantly lower in retail, as all these bad sticks will be repaired or filtered out in the factory before hitting retail stores.
We are going to be using this data to optimize the V12 release, currently under development.
Important qualification: The results might vary with different types of DDR5 RAM from different vendors built in different factories, or with different test hardware. But having some general indication is still a good thing.