Hardware issue found while doing DC

[kriesel]

Originally Posted by paulunderwood

Why not run version 7? Also why are you doing LL and not PRP?

V7.x won't do LL. He's repeatedly been advised to run PRP/GEC not LL, and continues to choose inferior reliability LL anyway. There are also known good interim PRP3 type 1 residues for known primes, and for both LL and PRP3 for large exponents.

Also, he hasn't yet indicated what model GPU he's trying to run on. No idea why that's still kept a secret while extremely unreliable.

On the same system (so I know it is not a difference in system side hardware or clocking or volts conditions, but one's on an extender & outside the case), I have two Radeon VII GPUs, both with Hynix VRAM, producing wildly different GEC error rates at same reduced power etc settings:
1 in 25.3 days = 0.04/day (~1G exponent, 56M fft, 17.03 bpw)
176 in 48.9 days = 3.6/day (and this one is running the less large exponent, 819751061, 48M fft, 16.29 bpw)
Almost two orders of magnitude ratio on error rate.
Operationally, they are almost equally productive, because the GEC saves the accuracy of the advancing iterations.

But at least one error per 3 minutes?! At least ~480./day, hundreds of times higher again. So frequently that progress can not be made with even LL's limited error checking. There's not just a hardware or configuration issue, it is MAJOR / FATAL.
With such extremely high error rate, gpuowl should be run with smaller -block and -log values than default.
Even if it does not begin to make net progress, it would give a better read of how high the error rate is.
See also https://mersenneforum.org/showpost.p...postcount=1131, https://mersenneforum.org/showpost.p...53&postcount=8 and the several posts following it.

[chalsall]

Originally Posted by joejoefla

Running DC of 62825017 on my cpu and GPU. Will report back on results.

Hey. Please forgive me for this, but I was watching out for the results of this test from your kit.

Did you run it to completion? Anything interesting? Your HW was the start of this whole conversation...

As previously documented, I don't get out much...

[joejoefla]

To kinda put this issue to bed, yes I've finally tested my RAM with memtest86. RAM was reporting issues at first. Then I started to test with one, two, then three sticks. I tested with four sticks but had to stop around the third test cycle because I had to use my desktop for work. No errors were reported then this whole time testing.

I'm going back and running a test on a known prime with my CPU downclocked to 4.0 from the default 4.9, with no undervolt. The motherboard is an Asus Z590. The RAM is 32GB of DDR4 2400Mhz. The CPU is running at 1.15v with a max temp of 70C. If I run I leave it at the stock 4.9, it'll throttle down to 4.6 and run at 92C.

Also, my GPU is a GTX 1080. Sorry for not putting it on there. It's not a secret.

[chalsall]

Originally Posted by joejoefla

To kinda put this issue to bed...

Thank you for the follow-up. Sincerely.

Few appreciate just how insane some kit can be.

[joejoefla]

It appears reseating the ram has "fixed the glitch". No more error messages. Tested in Memtest, ran some PRP cofactor checks, then ran a PRP double check with proof. Also ran a PRP on my GPU using GPUOWL 7. No more problems.