10900k álmost stable, except 384k FFT Prime 95

[rrul]

Hello,

After days (or actually weeks) of tweaking I have 5.2 steady and am just on the brink of 5.3.

It runs 30 passes of Linpack together with Youtube fine, small FFT Prime 95 fine, but as soon as a manually set 384k FFT size In-Place it errors. I raised the V very high (for some ridiculously so), to about 0.05 plus the Auto setting at x53 resulting in VID of 1.62 and Vout of 1.58 (LLC4 on MSI Unify), but as it rarely stays there for long with dynamic core speed (and throttling) I'm ok to risk the chance of degradation.

Thing is that even with these settings it crashes either quickly or after a bit, best run was 30 min, but a restart of that run crashed immediately.

L2 cache is 10x256kb and I might have this totally wrong but I guess that 384k sized FFT's will spill over into the 20mb L3 cache, so guess the instability is somewhere over there. Small FFT's that fit in L1/L2 cache run fine (without need for an offset).

I think I just hit a wall. For sure hit the temp limit as 3 cores throttle but dont mind that too much as irl in games I'll never hit the constant load that P95 brings so throttling there would be not much of an issue.

Tried LLC3 (almost flat), LLC5 with every combination of Vcore / SA / IO possible (latter two up to 1.4). Mem set at xmp 3600, 1.5 vdimm, seems stable and not the cause of the crashes.

Is there any setting besides Vcore I could change to stabilize the cache ?

So close, yet so far away


Thanks !

Ps. a blend run also crashed eventually, I think at 5k FFT size

 

[Woomack]

The main problem with these CPUs is that they're already maxed out and not much is helping in OC. Typical 10900K can make 5.4GHz on single cores, 5.1GHz on all cores with 300-400MHz offset for the cache. It gives ~100MHz average OC.
Since cache OC is not helping much on this CPU then I would set it lower and focus on something else.
Too high voltage makes everything worse as the CPU loses stability because of too much heat inside the chip. 1.55V is for sub-ambient cooling and you set more than that. Everything is integrated so some other components like a cache or memory controller can overheat and cause stability issues.

 

[EarthDog]

Why are you setting (what feels like) some arbitrary test in P95 and running it? You've run a couple of different tests that are stable, why not 'use' the PC for whatever it is for instead of chasing some odd instability down? If you're unstable in w/e you do, then maybe look at it.

+1 to Woomack.

 

[rrul]

Thanks Woomack. I indeed settled for 52 all cores. Really got carried away with voltages nearing 1.6.

By the way, on the 2x16gb 3600 C16 8Pack TG DDR4 we were talking about in the other thread. We both seemed to get 4200-16-17-38 stable at 1.5. It seems there was a little more left in the tank, 4300-16-17-38 1.5 seems stable as well. LP is running in the background as we speak and it passed 8 hours Aida already. Just tightened the subs a bit more, re-testing that now. If passed I will post confirmation in other thread.

Only annoying thing is I really aimed for sub 40 Aida latency and no matter which dial I turned I couldn't get it below 40.0 ns

Unless upping the Cache, but the cache on my cpu is super finnicky, it immediately becomes unstable at 4400, no matter SA/IO V's.

But after weeks of tuning I'm happy with the result. 10900k 5.2, 4300 C16.

Gives 635Gflops in 45 seconds in Linpack stress test nr. 5 and in Aida Read 64Gb Write 67Gb Copy 63Gb with low 40's latency.

Next problem to solve is WD Legion's benchmark; 3090 Strix runs 4k all maxed (incl RT) no DLSS around 35fps average, so that seems fine compared to others (although not really what I expected from a GPU that forces me to eat bread for a couple of weeks), but the benchmark 'impact' screen states that my fps is hurt by a low performing CPU

Maybe it only reads the default clock speed of 3.7Ghz (at least it displayes 10900k @ 3.7Ghz) but that would be pretty dumb for a benchmark.

The cpu seems to switch frequencies fine, in P95, Linpack etc HWinfo shows max speeds. TVB, Turbo 3.0, MSI Enhanced Turbo all disabled, just a regular All Core OC with Speedstep and Speed Shift still enabled.

Not sure what's up with that Ubisoft 'tip' on low performing CPU

 

[rrul]

Why are you setting (what feels like) some arbitrary test in P95 and running it? You've run a couple of different tests that are stable, why not 'use' the PC for whatever it is for instead of chasing some odd instability down? If you're unstable in w/e you do, then maybe look at it.

+1 to Woomack.

I homed in on 384k small FFT in place because that crashed the quickest. Small FFT runs fine, but 384k may need some more cache switching or something making it unstable.

Eventually FFT's like this pop up in P95 Blend as well so it doesn't pass a long run of Blend. And I do think any CPU should be able to pass that.

Fact that it passes Linpack no problem is probably due to AVX offset.

 

[EarthDog]

I homed in on 384k small FFT in place because that crashed the quickest. Small FFT runs fine, but 384k may need some more cache switching or something making it unstable.

Eventually FFT's like this pop up in P95 Blend as well so it doesn't pass a long run of Blend. And I do think any CPU should be able to pass that.

Fact that it passes Linpack no problem is probably due to AVX offset.

So... are you stable in the things you actually use your PC for or.........?

 

[BugFreak]

I say set it to 5.1 and run with it. If you start seeing crashes down the road then lower it otherwise let it run at 5.1 and save some wear and tear from the high voltages. I've had my overclocks crash after long stress test runs that never crash in day to day use that I call stable for years. Don't sweat one test crashing.

 

[rrul]

You have a point, but just call it OCD; I want the compu to be able to pass P95 Blend.

Stable on sight doesn't mean data is being corrupted in the background as far as I understood. Not every instability is visible via a crash.

But 5.2 seems to do be fine so settled for that. Got carried away with trying to hit 5.3.

 

[EarthDog]

You have a point, but just call it OCD; I want the compu to be able to pass P95 Blend.

Stable on sight doesn't mean data is being corrupted in the background as far as I understood. Not every instability is visible via a crash.

But 5.2 seems to do be fine so settled for that. Got carried away with trying to hit 5.3.

It doesn't mean there is any corruption at all. The point is if if isnt crashing and not borking data in the background (whatever that means) its stable for YOUR uses. People generally only use small fft for the cpu and blend for ram in p95. Setting custom lengths and running multiple other stress tests is living in the minutia for your stated purposes.

5.2 ghz and calling it a day sounds like a plan!