LinX - GFlops changes with vcore?

[wolfram23]

Hi all. Last night I was working on getting a stable 3.8ghz, 200x19. After I got the voltages high enough to be stable, I began reducing them back first vcore, at which point I was stable at 1.27V and then VTT which I got stable at 1.26V (If I remember, might be slightly off but the exact numbers aren't the point...). So I was running Prime95 for quick stability checks, since having the voltages too low I find it'll stop a core or two within the first 3 tests. Then I used LinX for a little bit more checking, plus it's nice to have a speed output.

I had an interesting result. With only changing vcore, I had it at 1.27 and got around 55.5 GFlops on all 5 passes, no errors (as high as 55.8 I think). The theoretical max is (16x3.8) 60.8, so it's only 8.7% off. Really good imo. I then dropped the vcore to 1.26V. Stable in Prime95, moved on to LinX. Almost every pass was in the 51-52 range if stable, and one time it did hit the 55GFlops but it was on a first pass and stopped from an error. I made sure all the apps were closed and processes stopped so it was in the same test conditions. Is this reduction in speed because of throttling perhaps? After trying a few passes with similar results, I went back up to 1.27V vcore and again got stable results at 55.5GFlops.

So can anyone explain what happened? I can't rule out the possibility of background processes but I really don't think that was the issue. So was it throttling? Or maybe there's some "magic" that happens when you have just the right voltages...

Just as a point of interest, I spent a little less time tweaking the voltages at 3.6ghz (200x18) - 1.18V Vcore and 1.17V VTT - and LinX gives me around 9-10% off theoretical max, so I seem to be more efficient at 3.8ghz. I've also read that an odd multiplier gives better results tho so maybe that's why? As in x19 is better than x18.

 

[Scotteq]

I've heard the "Odd Multi" thing from a number of places. Anecdotal evidence for sure, but I'm hearing it often enough that it's certainly in my mind when I play with my own rig. Having said that, my 960 is currently at 24 so...

Another thing I've heard, and can say from personal experience appears to be true, is that LGA1366 likes a little extra Uncore voltage as the bclk rises. And that doing so lets you get away with running a little less vCore. Since your rig passes at least the preliminary tests in P95, you might want to bump the uncore a notch at the lower (1.26v) vCore setting and see does that make a difference.

Regarding the differences in throughput, I suppose if the system were right on the margin it could have a little trouble maintaining it's speed while not necessarily throwing a detectable error... But I haven't played with LinX enough to have more than a vague supposition on that.

I'd be interested in hearing if the Uncore suggestion has any effect on throughput, as well. I may have to go play with it myself over the weekend.

 

[wolfram23]

Hmm interesting suggestion. I'll give it a try tonight and see what happens, tho I am fairly satisfied with the current settings. Definitely worth playing with haha.

And actually I ran Prime95 for about 3.5hrs when I went to bed last night without errors. I know it should be longer for "true stability" but that's how long it took to go to sleep and stop hah.