Hi, Philip, All these statistical analyses assume that errors are not correlated in many dimensions. If that were to happen on the wire, it would likely be caught by things like Ethernet’s CRC. The issue we’re dealing with are errors that are either not caught on the wire OR happen in memory or processing, neither of which are very likely to be uncorrelated. The only thing that’s “pretty clear” to me is that if this were actually as uncorrelated as you’re assuming, we would have lots of examples of transfers that failed. I asked a simple question - are we seeing that? I’ll admit that doesn’t come up for me, but by everyone’s “analysis” they ought to be raining down everywhere, esp. in large data centers. But I haven’t heard anyone screaming yet. So either the other protections we have (as others have mentioned) are sufficient or these errors are not correlated the way that’s been assumed (and thus they’re being caught by the existing checksum). There’s no law against repeating a checksum with different data if all the data gets there OK. Joe
