Hardware errors have occurred during the test!

[retina]

Originally Posted by drkirkby

I believe that some non-ECC RAM has a parity bit, so can detect errors. But perhaps it is rare.

It is so rare that using only a parity bit now only exists in machines stored in museums.

ECC is a logical extension of parity anyway, so maybe you are confusing some naming where someone refers to ECC as being multiple parity bits? Technically it is just multiple parity bits I suppose, but to call it parity is not giving it its proper due, and would be misleading.

[drkirkby]

Originally Posted by retina

It is so rare that using only a parity bit now only exists in machines stored in museums.

Yes, I am probably thinking of older PCs. I am fairly certain that some PCs, going back to the 80286/80386 era, had a parity bit for error detection, but not correction. But maybe I am mistaken. Vg

Originally Posted by retina

ECC is a logical extension of parity anyway, so maybe you are confusing some naming where someone refers to ECC as being multiple parity bits? Technically it is just multiple parity bits I suppose, but to call it parity is not giving it its proper due, and would be misleading.

No, I was not confusing them. Wikipedia says
“Most non-ECC memory cannot detect errors, although some non-ECC memory with parity support allows detection but not correction.”
That is what I was thinking of - RAM with a single parity bit, but not ECC RAM.

Unfortunately ECC RAM is considerably more expensive than non-ECC RAM. I am guessing that it is a smaller market, although with a lot of cloud computing services around, I would expect the servers to be using ECC RAM, so maybe the relative cost between ECC-RAM vs non-ECC RAM might fall. If one is running Windoze, the reliability of the the OS doesn’t warrant using ECC RAM.

[kriesel]

IIRC some of the later 3rd party memory addin cards for the original IBM PC 8-bit-wide data bus offered parity.

Slightly later, FPM with parity https://www.ebay.com/itm/153940233188
1986, for 486 cpus https://en.wikipedia.org/wiki/Fast_Page_Mode_DRAM

ECC or bust seems the norm nowadays or even in workstations bought used a few years ago. Mere parity checking is not seen on offer these days.
https://en.wikipedia.org/wiki/RAM_parity

[R. Gerbicz]

Originally Posted by drkirkby

Depending on your operating system, and hardware, errors may be logged. With the exception of my laptop, all computers I use have error correcting (ECC) RAM. With that, most RAM errors get logged, and usually corrected, so the application doesn’t know about it. I think even standard RAM will detect errors, although not correct them. This might be logged. If you see errors about the same DIMM or same CPU, it would be wise to replace it, although it could be a motherboard fault.

That expensive ECC memory doesn't detect (computational) FFT errors when you're using floating point FFT. My check detects those errors also, so you don't need to use suboptimal larger FFT size or pure integer FFT.

[Zhangrc]

Originally Posted by MarkVanCoutren

I've been having similar issues. I haven't tried to overclock but I left it running for a few days.

Iteration: 8280000 / 108671053 [7.61%], ms/iter: 10.106, ETA: 11d 17:48
Hardware errors have occurred during the test!
1 Gerbicz/double-check error.
Confidence in final result is excellent

I have an Intel Core i5-9600K 3.7 GHz 6-Core Processor running Windows 10. I think I got it on my last number too. Should I just let it finish the number and see if it goes away?

I suppose it's related to roundoff errors.
You are using FFT length 5760K, right? For your exponent (about 108.7M), the FFT length might not be enough and the roundoff will go too high (say, >0.4). A 6M fft will be sufficient.

[VBCurtis]

Originally Posted by Zhangrc

I suppose it's related to roundoff errors.
You are using FFT length 5760K, right? For your exponent (about 108.7M), the FFT length might not be enough and the roundoff will go too high (say, >0.4). A 6M fft will be sufficient.

No, Prime95 does not call that a hardware error. Prime95 retries the computation when roundoff is too big, and if reproducible it prints a "error is reproducible, not a hardware error." Usually (always?) it also bumps up the FFT and continues.