Hacker News Reader

Comment by anonymousiam | Hacker News Reader

I am surprised that ECC is not more popular. RAM errors (SEUs) are common in every platform. One hacker that I worked with almost a decade ago did this BlackHat presentation with some analysis that correlates errors with increased temperature (by geo-locating clients of "bit squatting").

I put together a new SuperMicro server a few months ago and went with 256GB of ECC. Yeah it's very expensive, but it's absolutely worth it if you care about the integrity of your data.

See: https://media.blackhat.com/bh-us-11/Dinaburg/BH_US_11_Dinabu...

Where are you now Artem?

View on HN

It's not even that expensive. An 16GB "server" DIMM costs what, $65 at retail? Apple will charge you $200 for 8GB of non-ECC. The real cost on ECC memory is getting a CPU and platform that support it properly.

> An 16GB "server" DIMM costs what, $65 at retail?

Clocked at 2400MHz, sure. I can't find a single non-sketchy listing for even 3200MHz unbuffered ECC memory, let alone the 3600MHz you want for current Ryzen chips. And those are just the speeds at reasonable prices, ignoring the 4000-5000 range.

At least DDR5 should improve things.

> I can't find a single non-sketchy listing for even 3200MHz unbuffered ECC memory

https://www.crucial.com/memory/server-ddr4/mta18adf2g72az-3g...

> let alone the 3600MHz you want for current Ryzen chips

Current DDR4 JEDEC-approved speeds top out at DDR4-3200. Anything beyond that is technically overclocking, even if it's overclocked from the factory and marketed to run at those speeds. Assuming that your platform supports overclocking DRAM, you can also run DDR4-3200 ECC memory at 3600 MT/s, if you feel that the performance uplift is worth the risk of stability issues.

And the good thing is that you’ll know immediately about stability issues with ECC.

Patrol scrubbing is a RAS feature of only more expensive platforms. On most hosts you are likely to learn about the issue only at the time some program tries to access the marginal page or row.

> Patrol scrubbing is a RAS feature of only more expensive platforms.

All of the ECC-capable platforms I've used (including my AMD-based desktop PCs) have supported patrol scrubbing. What have you used that doesn't?

My AM4 motherboard doesn't offer any options for scrubbing and its manual is vague about whether ECC errors will even be reported to the operating system. By contrast my HPE ProLiant DL360 has firmware settings for scrubbing and I know the event reporting works because I see it in the logs all the time.

The operating system will explicitly tell you at runtime if ECC support is available. If this returns an affirmative value, then ECC is working.

Support for patrol scrubbing is a bit more variable. It's supported in my ASRock motherboard (B450M Pro4). It was also supported in two previous ASUS motherboards (AMD AM2 and AM3 platforms, respectively), although they called it something else that I can't remember. As well, a lack of a setting for patrol scrubbing doesn't necessarily mean that it's unavailable; it could just be enabled but not user-configurable.

This is OS-configurable. Look for /sys/devices/system/edac/mc/mc*/sdram_scrub_rate on Linux.

The motherboard doesn't really get a say on whether the OS can correctly use ECC support on AMD as far as I know. As long as the memory is correctly initialized in ECC mode, the OS should always be able to use its EDAC driver to interact with it.

> The motherboard doesn't really get a say on whether the OS can correctly use ECC support on AMD as far as I know.

This isn't entirely true. The motherboard needs to physically support ECC (i.e., traces for a 72-bit memory bus need to be present) for it to work. ECC support can also be disabled within the BIOS firmware, either explicitly via a user-adjustable toggle, or by the motherboard vendor hard-coding the setting to disabled.

In any case, if the OS reports ECC as working, then ECC is fully functional.

My point is that once RAM is initialized in ECC mode, then the OS will see it. The motherboard can't support ECC mode but then screw up EDAC. That's why I said "as long as the memory is correctly initialized in ECC mode".

It's a binary thing. Either the motherboard supports ECC, or it doesn't. If it supports ECC then you can always tweak the scrubbing from the OS.

Some people claim Gigabyte X570 motherboards in theory have the traces needed for UDIMM ECC, but the (hardcoded) BIOS settings cause the memory to not initialize in ECC mode.

Good, glad it exists, though it looks like it's only been available for about a year.

I don't care about JEDEC. I care about the XMP spec that's promised to me by the manufacturer, so I can rely on it. I'm not against overclocking myself but if I'm trying to boost it by a significant amount then I have no idea what to set the timings to and there's a good chance it just won't work, with no recourse on my end.

Also that module costs way more than $65.

I thought we were talking about correctness, not going shopping for more-error-prone DIMMs.

Well I thought we were talking about getting ECC without major tradeoffs, to understand its [lack of] popularity.

And I wouldn't say faster memory is necessarily more error prone. But it doesn't really matter, because non-ECC products have to have extremely low error rates to be viable. Even if a speed is at the high end of "extremely low", once you add ECC you'll have an exceptionally solid component.

> Well I thought we were talking about getting ECC without major tradeoffs, to understand its [lack of] popularity.

ECC isn't popular on desktops because Intel, which supplies the overwhelming majority of the processors for platforms that use DIMMs, doesn't support ECC at all. As such, the market for people who can actually use ECC in a desktop (much less would want to) is very small. The only reason unbuffered ECC DIMMs are even on the market at all is because low-end servers use them.

I'm not sure where everyone gets the idea that Intel charges a big premium for ECC platforms. If you buy the low-end workstation Xeons the processor costs about the same. For example, Xeon W-1290 (10C, 5.2GHz turbo) has a $494 MSRP. The i9-10850K (also 10C, 5.2GHz turbo) has a $453 MSRP. You also have to buy a motherboard with a workstation chipset to use the Xeon, but W480 chipset motherboards are also priced about the same (~$200) as their consumer Z490 counterparts.

The real reason ECC is not common is because it is perceived as unnecessary for desktop workloads. Software crashes and corrupts data all the time because of bugs, not bitflips. Most consumers care so little about their data that a single hard drive crash will wipe it all out anyways. Why spend 12% extra on memory if consumers hardly see any value from it?

Four of these weren't cheap. ($1,441.36)

https://www.amazon.com/gp/product/B07YF249FX

Reply on news.ycombinator.com