Hacker News Reader

Comment by IceyEC | Hacker News Reader

Any word on ECC support for the Ryzen chips?

I thought ECC was already supported out of the box on Ryzen? IIRC, the issue was motherboard support (motherboards have been the pain point for Ryzen from the very start).

http://www.hardwarecanucks.com/forum/hardware-canucks-review...

As a current user of ECC RAM w/ a Ryzen 7 1800, I can tell you it works.

We need faster ECC RAM though: the external RAM speed affects the internal AMD Infinity Fabric, which provides inter core comms.

It works but does it "work"? I've never seen a straight answer on this. It's not validated - does it really work, as in, error corrects?

I don't remember the source, but one of the review sites overclocked some ECC ram for Ryzen to the point where it was erroring. They were stable and viewing the one-bit errors coming in and being corrected until they got a two-bit error and the machine booted(as they are supposed to).

No, two-bit failures are supposed to hard fault (to minimize damage of data corruption). Which the Ryzen did not do (in that specific test, at least).

The OS detected the fault and terminated the application which used the faulted memory address range, which is what it was supposed to do.

The guy who ran the exact test OP is referring to specifically states that it handles two-bit errors in a sub-optimal manner.

Sure, you don't have to hard fault. Just like regular memory controllers don't fail on any bit-errors. But it's ideal, when dealing with critical data.

Memory (and other) errors generate a Machine Check Exception which are handled by OS and/or BIOS. OSs can be smart enough to accumulate a history of errors and offline affected physical pages (which can result in a very localized single process segmentation violation). You could just log the errors. But you could also dial up the reaction all the way to NMI on these errors, which often triggers a processor reset.

It definitely works on ASRock AM4 boards. Just make sure to get unbuffered ECC RAM.

Ryzen will work with ECC ram but it won't actively correct ECC errors (at least on Linux) the interrupt code flags that as differed.

Currently there aren't Ryzen server chips out, just desktop/enthusiast models.

And word on why this is, or a source for it? I'm curious if this is just something that can be fixed with an upgraded kernel.

Yes, it does. It'll correct 1-bit errors and report 2-bit errors. It just doesn't fault on 2-bit errors, as is expected.

Ryzen will send a machine check exception to the OS in the event of a double-bit error. It's up to the OS on how to handle it.

This also isn't a Ryzen-specific behavior. All x86 ECC-capable machines work this way.

4.10.something (4.10.6 maybe?) Added EDAC support for ECC.

Reply on news.ycombinator.com