16 ms·

I feel the need for ECC is a bit overstated, probably confused by listening too much to people in the server space. In servers, ECC is critical because servers

I feel the need for ECC is a bit overstated, probably confused by listening too much to people in the server space. In servers, ECC is critical because servers simply have more RAM than PCs do. You can easily find systems with 32 sticks of RAM. If you roll 32 dice, odds of a rare freak error is a lot bigger than if you roll 2. It's the same reason why you probably don't need RAID on a computer with one or two disks. Stick 48 disks in a machine however, and you're a fool to go without it (or a plan at least). It's why you don't need an automatic fire suppression system in or homelab, because the odds of a fire in one or a few PSUs is fairly small, but stick a few thousands in a room and suddenly it's a very real problem. That, and most memory problems in PCs tend to be from dodgy overclocking or just bad sticks rather than cosmic rays. ECC won't really save you from that.

Vogtinator 3y ago
Counterpoint: Without ECC, you simply don't notice how many flips happened and that they were the cause of some strange behaviour.
- marginalia_nu 3y ago
  Most of them simply don't matter though, and are inconsequential along other system instabilities. It's just neurotic to worry about issues when you couldn't tell they existed without being informed of them.
  - ilyt 3y ago
    Most issues with car don't matter for driving, till it catches on fire or crashes due to catastrophic failure. Why ever go to mechanic ?
    - marginalia_nu 3y ago
      Most car crashes just aren't due to mechanical failure.
      - ilyt 3y ago
        It would be if you had anything to do with designing or servicing them...
Modified3019 3y ago
>dodgy overclocking Error detection and correction is actually ideal when overclocking. It'll not only save you, it'll let you know you need to back off a bit, because the handful of errors that may happen once a month despite passing days of tests will show up in the system logs instead of staying silent and causing problems. I speak from experience as I overclocked ECC memory in a first gen threadripper system, playing around with overclocking as a hobby for a bit. ECC memory was fantastic to work with. However the dodgy or outright lack of CPU/Motherboard support results in trying to market that feature to gamers/overclockers having too much friction. Much like what went on with intel's confusing proprietary optane/flash combo M.2 drives that needed certain motherboards in order to fully access both parts of the drive. And with bad sticks, instead of hard to explain crashes, you'll end up with system logs filled with corrected/detected errors and maby crashes if it's really bad. Then you just RMA the sticks like normal. So honestly that's a win too, because that's the whole point of detecting and correcting errors. It's in every other part of the system already, continuing to leave this capability out of memory is an oversight that needs to be.. corrected. I just don't really get the strange resistance people have to ECC, it's not some sacred cow. It's just good sense.
ilyt 3y ago
> It's the same reason why you probably don't need RAID on a computer with one or two disks No, it is not "the same reason" Disks have ECC, even consumer hard drives have checksummed blocks, that's how you get media errors detected. RAM does not. Well, technically with DDR5 it has internal one but it does not give any feedback to the machine so you might not know your RAM has any problems. > It's why you don't need an automatic fire suppression system in or homelab But you want smoke sensors. ECC is the smoke sensor too > That, and most memory problems in PCs tend to be from dodgy overclocking or just bad sticks rather than cosmic rays. ECC won't really save you from that. But it will tell you the problem Very ignorant take
- marginalia_nu 3y ago
  > Disks have ECC, even consumer hard drives have checksummed blocks, that's how you get media errors detected. RAM does not. Well, technically with DDR5 it has internal one but it does not give any feedback to the machine so you might not know your RAM has any problems. Sure, and disks have random errors even with error correction. Adding hamming codes just makes errors less likely and easier to detect, but they aren't fool proof, in ram or on disk or anywhere else. The protection offered is a bit questionable because they rely on the very sketchy assumption that errors are statistically independent events, which is rarely how storage errors work. > But it will tell you the problem > Very ignorant take Hardware isn't perfect. No matter how much error correction you pile onto it, you will still have errors and some cases will still be undetectable. It's a matter of pushing the rate of errors into an acceptable range, as a pure cost-benefit tradeoff, statistics through and through.

Vogtinator 3y ago

Counterpoint: Without ECC, you simply don't notice how many flips happened and that they were the cause of some strange behaviour.

marginalia_nu 3y ago
Most of them simply don't matter though, and are inconsequential along other system instabilities. It's just neurotic to worry about issues when you couldn't tell they existed without being informed of them.
- ilyt 3y ago
  Most issues with car don't matter for driving, till it catches on fire or crashes due to catastrophic failure. Why ever go to mechanic ?
  - marginalia_nu 3y ago
    Most car crashes just aren't due to mechanical failure.
    - ilyt 3y ago
      It would be if you had anything to do with designing or servicing them...

Modified3019 3y ago

>dodgy overclocking Error detection and correction is actually ideal when overclocking. It'll not only save you, it'll let you know you need to back off a bit, because the handful of errors that may happen once a month despite passing days of tests will show up in the system logs instead of staying silent and causing problems. I speak from experience as I overclocked ECC memory in a first gen threadripper system, playing around with overclocking as a hobby for a bit. ECC memory was fantastic to work with. However the dodgy or outright lack of CPU/Motherboard support results in trying to market that feature to gamers/overclockers having too much friction. Much like what went on with intel's confusing proprietary optane/flash combo M.2 drives that needed certain motherboards in order to fully access both parts of the drive. And with bad sticks, instead of hard to explain crashes, you'll end up with system logs filled with corrected/detected errors and maby crashes if it's really bad. Then you just RMA the sticks like normal. So honestly that's a win too, because that's the whole point of detecting and correcting errors. It's in every other part of the system already, continuing to leave this capability out of memory is an oversight that needs to be.. corrected. I just don't really get the strange resistance people have to ECC, it's not some sacred cow. It's just good sense.

ilyt 3y ago

> It's the same reason why you probably don't need RAID on a computer with one or two disks No, it is not "the same reason" Disks have ECC, even consumer hard drives have checksummed blocks, that's how you get media errors detected. RAM does not. Well, technically with DDR5 it has internal one but it does not give any feedback to the machine so you might not know your RAM has any problems. > It's why you don't need an automatic fire suppression system in or homelab But you want smoke sensors. ECC is the smoke sensor too > That, and most memory problems in PCs tend to be from dodgy overclocking or just bad sticks rather than cosmic rays. ECC won't really save you from that. But it will tell you the problem Very ignorant take

marginalia_nu 3y ago
> Disks have ECC, even consumer hard drives have checksummed blocks, that's how you get media errors detected. RAM does not. Well, technically with DDR5 it has internal one but it does not give any feedback to the machine so you might not know your RAM has any problems. Sure, and disks have random errors even with error correction. Adding hamming codes just makes errors less likely and easier to detect, but they aren't fool proof, in ram or on disk or anywhere else. The protection offered is a bit questionable because they rely on the very sketchy assumption that errors are statistically independent events, which is rarely how storage errors work. > But it will tell you the problem > Very ignorant take Hardware isn't perfect. No matter how much error correction you pile onto it, you will still have errors and some cases will still be undetectable. It's a matter of pushing the rate of errors into an acceptable range, as a pure cost-benefit tradeoff, statistics through and through.