I am one of the principle engineers working on a game that was able to support over 1.2 million players concurrently, so I feel that I have some merit behind what I'm about to say.
ECC Memory does not have such an extreme overhead. Our gameserver is memory bandwidth bound. To put that into perspective a change from DDR3 to DDR4 saw a jump in performance of 34%~. Using non-ecc memory (which is obviously very available to the development team) only gave a penalty of roughly 3% in the worst case. (both DDR4 systems).
Why are we memory bandwidth bound? because we allocate 250GB of "game world" and phases and make teeny tiny modifications to them when you run around or pick up loot and shit. It's as taxing as you can be on memory bandwidth since you can't batch set or load anything.
On a further note, while the implementation of ECC itself is very fast for modern RAM, the same is not true of VRAM on GPUs. At least as implemented in NVidia Tesla GPUs, enabling ECC has a huge performance penalty by itself. [1]
[1] http://geoceano.fr/wp-content/uploads/2015/01/cpu-gpu-lbm-in...
When you start looking at multi CPU computers with 1+TB of ram the story changes.
PS: Some of this stuff is also really hard to test. Without ECC it's hard to data in terms of stability for petabytes of ram over years. And no extrapolations don't work very well.
Now, non ECC memory might have better specs / a higher overclock rate etc, but doing either of those increases the rate of errors. And sure you might not notice the problem directly, but that does not mean things are ok.
The value of ECC is often more to detect errors than to correct them. Remember, if you have 10 errors an hour then you unlikely to see them in a 20 second memory test.
In practice out of 10,000 machines you might have 1/2 your errors on 10 of them. Critically, if you have ECC memory and you both detect and replace those 10 machines the apparent rate of failures then drops.
PS: This is a little dated (09) and only from a single set. But, still an interesting introduction: https://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf "Around 20% of DIMMs in Platform A and B are affected by correctable errors per year, compared to less than 4% of DIMMs in Platform C and D. ... About a third of all machines in the fleet experience at least one memory error per year (see column CE Incid. %) and the average number of correctable errors per year is over 22,000. "
EDIT: the least out of date reference I could find is a blog post by Jeff Atwood, where he concludes that based on the available literature ECC-correctable errors are quite rare. [1]
"The median number of errors per year for those machines that experience at least one error ranges from 25 to 611." "We find that for all platforms, 20% of the machines with errors make up more than 90% of all observed errors for that platform."
A minority of that 20% again have even more issues.
The worse case is literally 10,000x more than 1 bit error per year on a single machine. Without using ECC you really can't tell other than some vague not sure what's wrong just replace it.
In terms of speed it's ~2–3 percent slower worst case, and memory is only one component of overall performance lowering the impact further.
More importantly, bit errors are not evenly distributed with some users seeing vastly more errors.