Every dedicated server quote has a RAM line, and most buyers skim past it. Then one quote says ECC and the next says non-ECC at the same capacity for less money, and the decision gets made on price. That is the wrong cut. ECC is a real feature with a real cost, and whether it is worth paying for depends almost entirely on what the box does between reboots. Here is the plain version.
ECC stands for error correcting code. The memory controller stores extra bits alongside your data and uses them to detect and fix errors on the fly. Most server memory uses single-error correct, double-error detect, shortened to SECDED: a single flipped bit in a word gets corrected silently, and two flipped bits get detected and reported rather than served to your application as good data. Non-ECC memory has no such check. If a bit flips, the flipped value is what your program reads, and the corruption moves downstream until something crashes or, worse, something writes it to disk.
The natural question is how often a bit actually flips. There is published data on this. The large fleet studies are the reference point: Google's 2009 paper on DRAM errors in production and the 2015 study from Facebook's fleet both measured correctable errors on real hardware at scale, and their headline finding is the useful one. Roughly a few percent of DIMMs log at least one correctable error per year, and once a DIMM has produced errors, its error rate stays elevated, which points to the module rather than pure chance. A single server sits at the small end of that distribution, but a single server also runs for years. Over a three year rental the expected number of correctable events on a non-ECC box may still be small. The tail of that distribution is what you are insuring against: an uncorrectable flip in a database page, in a filesystem structure, or in a long-lived in-memory index.
Any workload where a wrong bit survives longer than a process restart:
The pattern is memory holding state across time. That is what ECC protects. If state can be rebuilt from disk in seconds after a crash, the insurance matters less.
Static and cache-heavy web serving is the classic case. A WordPress or nginx box serving cached pages rebuilds its working set from disk on every restart, a restart takes seconds, and the page cache can be dropped and refilled at any time without the visitor noticing. CI runners and build agents are similar: a corrupted build fails a test, the runner restarts, the next build is correct. Game servers and stateless APIs sit in between, since a crash drops sessions but recovers on restart.
None of this means those boxes never need ECC. It means a crash there is an inconvenience you can measure in seconds of restart, while a silent write-back corruption in a database is the kind of problem you find weeks later in a backup restore. Price the two outcomes differently.
Four things hide in that line, and each changes the price:
At the time of writing, a 32 GB ECC UDIMM typically costs a bit more than a same-spec non-ECC DIMM, and the real jump comes when you move from consumer platforms to server platforms: an ECC-capable board and CPU often costs meaningfully more than the consumer equivalent, and that difference lands in the monthly rental. The usual rough figure we see in quotes is a 10 to 20 percent premium for an ECC platform at small capacities, narrowing as capacity grows, because RDIMM capacity is cheap per gigabyte once you are on a server platform anyway. Treat those as ballparks, current pricing moves month to month.
There is also an availability argument that no price sheet shows. Server boards, server DIMMs and IPMI boards come from the same ecosystem, and that ecosystem assumes ECC. The box that boots a remote recovery console when the kernel panics is described in our piece on out-of-band access, and it is a server board, which almost always means ECC is already there.
If the machine holds state that would take more than a minute to rebuild, buy ECC. If the machine can be power-cycled into a correct state, spend the difference on capacity, NVMe or a second provider. Most of the dedicated boxes we quote sit in the first group, which is why ECC is the default on our configurations. If you want a recommendation for a specific workload, mail admin@servercabin.com with what runs on the box and how it restarts, and we will quote both ways.