3 ms·

> I have a Ryzen desktop with ECC, and it registers about one bit-flip per week. This isn’t normal. You have a bad memory module. A non-ECC machine should abl

> I have a Ryzen desktop with ECC, and it registers about one bit-flip per week. This isn’t normal. You have a bad memory module. A non-ECC machine should able to support memtest86 (a rigorous memory testing tool) for a week straight without a single bitflip. Having a bit flip per week is so far away from normal that it’s definitely a bad memory part.

dijit 4y ago
> > I have a Ryzen desktop with ECC, and it registers about one bit-flip per week. > This isn’t normal. You have a bad memory module. > Having a bit flip per week is so far away from normal that it’s definitely a bad memory part. It's hard to know what's normal. If you're somewhere in the world like northern Sweden the chance of cosmic rays goes up from being less protected from the suns rays (Hey, but northern lights!) and if you're at higher altitude; along with less interesting things such as how shielded the ram is (server chassis, if you removed the rack doors or not). Memory density has a lot to go with it too, the chances of bitflips increases a lot. Anyway; regardless of the theory behind it: I've run global operations with multiple-thousands of physical machines with high density ram (and 16-32 DIMMS in each server) and noticed at least 1 bit flip a week on average for each machine. But some locations were worse impacted than others. Annoyingly, ECC is not created equal, there are some correctable bit-flips that are hidden from the iDRAC/iLO: this is the same type of ECC that the rPI is using (referred to as on-die): https://datasheets.raspberrypi.com/rpi4/raspberry-pi-4-product-brief.pdf https://datasheets.raspberrypi.com/rpi4/raspberry-pi-4-produ... As with all things like this YMMV, but ECC is doing a lot of heavy lifting.
moloch-hai 4y ago
Memtest86 will only notice a bit-flip by sheer luck: the flip has to happen between when it wrote and when it reads back a short time later. Any flip in bits it is not immediately looking at will slip by. Memtest86 is not magic. It is designed to look for systematic failures, even if intermittent. It does that pretty well given it has only bus-level access.
Gordonjcp 4y ago
I'd hate to think what all those neutrinos would be doing to you. You are very much larger than a memory cell, for every bit flip you must be catching a trillion neutrinos. Maybe OP's house is in the Van Allen Belt or something?

dijit 4y ago

> > I have a Ryzen desktop with ECC, and it registers about one bit-flip per week. > This isn’t normal. You have a bad memory module. > Having a bit flip per week is so far away from normal that it’s definitely a bad memory part. It's hard to know what's normal. If you're somewhere in the world like northern Sweden the chance of cosmic rays goes up from being less protected from the suns rays (Hey, but northern lights!) and if you're at higher altitude; along with less interesting things such as how shielded the ram is (server chassis, if you removed the rack doors or not). Memory density has a lot to go with it too, the chances of bitflips increases a lot. Anyway; regardless of the theory behind it: I've run global operations with multiple-thousands of physical machines with high density ram (and 16-32 DIMMS in each server) and noticed at least 1 bit flip a week on average for each machine. But some locations were worse impacted than others. Annoyingly, ECC is not created equal, there are some correctable bit-flips that are hidden from the iDRAC/iLO: this is the same type of ECC that the rPI is using (referred to as on-die): https://datasheets.raspberrypi.com/rpi4/raspberry-pi-4-product-brief.pdf https://datasheets.raspberrypi.com/rpi4/raspberry-pi-4-produ... As with all things like this YMMV, but ECC is doing a lot of heavy lifting.

moloch-hai 4y ago

Memtest86 will only notice a bit-flip by sheer luck: the flip has to happen between when it wrote and when it reads back a short time later. Any flip in bits it is not immediately looking at will slip by. Memtest86 is not magic. It is designed to look for systematic failures, even if intermittent. It does that pretty well given it has only bus-level access.

Gordonjcp 4y ago

I'd hate to think what all those neutrinos would be doing to you. You are very much larger than a memory cell, for every bit flip you must be catching a trillion neutrinos. Maybe OP's house is in the Van Allen Belt or something?