I wonder if semi-reliable RAM could be made to work for training. After all gradient descent already works in a stochastic environment, so maybe the noise from a few flipped bits doesn't matter too much.

Also, depends on the nature of the error. If only a small memory range is affected, you could patch the kernel to avoid it.

No need for patching you can disable specific ranges on Linux using the memmap kernel parameter. It's often used for that purpose.

Sorry but I am looking for reliability. Do you have any stick that fails 100% of the time?

Have you tried underclocking or loosening the timings?

my experience is if even memtest86 fails the memory is truly borked.

Well, it kind of depends. With XMP (which is overclocking) I've found plenty of kits on Ryzen not passing memtest with the XMP settings. Different CPUs seem to be able to run their memory controller harder without error.

And then there are other factors like more sticks of ram stressing things further. I had to downclock to get memtest stable when running 4 sticks even though each kit ran fine on it's own. But that is expected as well as 4 sticks stresses the memory controller even further.

I confess I don't have any real recent experience with DDR5 though, mostly with DDR4 on Ryzen 1000-5000 series.