Similar to CPUs, where many arrays have spare yield capacity, even whole cores can get disabled (and possibly sold in a different bin). DRAM stores redundant electrons in capacitors to patch it up and boost yields. Everything in reliability is a spectrum.

"ECC" does not give you fully reliable RAM. UEs are still be observed.

What's the chance of fail? If you have one device that achieves equal performance with less reliable cells and redundancy to another device that uses more reliable cells without redundancy, it's not really any different.

NAND is horribly flaky, cell errors are a matter of course. You could buy boutique NOR or SLC NAND or something if you want really good cells. You wouldn't though, because it would be ruinously expensive, but also it would not really give you a result that an SSD with ECC can't achieve.

The net error rate is lower with the internal ECC.

DDR4 is not fully reliable memory either.

This is common for many high speed electrical engineering challenges: Running a slightly higher error rate option with ECC on top can have an overall lower error rate at higher throughput than the alternative of running it slow enough to push the error rate down below some threshold.

It makes some people nervous because they don’t like the idea of errors being corrected, but the system designers are looking at overall error rates. The ECC is included in the system’s operation so it isn’t something that is worthwhile to separate out.