Playing with my ring

[mackerel]

Ring cache that is.

System is 6700k based. I'm aiming to optimise Prime95-like performance and don't care about anything else. This can be broken down into two test conditions: those workloads that fit inside L3 cache, and those that don't. In the latter case, ram is usually limiting. Putting that aside, the L3 cache isn't unlimited either. So I'm wondering if OC'ing that might help. Typically I leave it on auto, which on the mobo I use is 4 GHz. To my understanding, the L3 cache = ring cache.

I've done a slight core overclock so far, with 4.2GHz 1.25V seeming ok. Actually better than stock, as the system does 4.0 GHz at 1.33V+ for some stupid reason.

Next I tried jumping to 4.3 GHz for both core and ring. Booted ok, locked up on P95 stress. Not good. Put ring back to auto (4 GHz) and it has been fine for a while. I'm typing this over lunch, so will go back and play some more after that.

Is there any info on OC'ing the ring? A lot of info out there is gaming biased so useless. One place suggested 42x was easy but you have to pile on the volts to go above that, which is not inconsistent with what I've seen so far.

Thought does occur, I could test it backwards... if I lower the ring clock, do I see a drop in performance? If not, then I'm unlikely to see a gain from raising it.

 

[mackerel]

Done some quick testing now. Testing was 4x 256k FFT, which in total is about 8MB dataset, or equal to the L3 cache (excluding code and other overheads).

CPU was fixed at 4.3 GHz. Ring tested at 3.5, 4.0, 4.2 GHz. Normalising to 4.0 GHz as reference, 3.5 was 2.6% slower, 4.2 was 0.4% faster. So the difference is not zero, but it isn't significant either.

Next is to see how core clock does... past testing suggested it scaled near perfectly so this should provide bigger gains.

 

[Kenrou]

I managed mine at 4.7ghz/4.7ghz 1.33v the only difference I saw was in Aida64 memory tests, gaming-wise was maybe 1fps gain ? Possibly not even that ?

 

[mackerel]

I have a self imposed limit of set 1.25v for power limiting (I run 24/7 so it adds up). Long story short, not even 4.2/4.2 was stable at that voltage. I'm back testing at 4.2/4.0 now. Maybe 4.1 ring would be ok, but I've reached the point where I don't care any more about that 0.2%.

 

[Kenrou]

Stock ring is 4.1, and the several reviews I saw so far told me that ring should be set around 200mhz below clock for optimal stability (discussed it with E_D and several others on my own 6700k thread if memory serves). If I leave it on auto for example and bump clock to 4.8ghz ring will rise to 4.5/4.6 depending on voltage.

 

[mackerel]

If I fix the GPU to 4.2 and leave ring on auto, it stops at 4.0. It does look like the ring itself might need more voltage as you turn it up, like the core. So if OCing the core, the increased voltage from that also feeds the ring. So taking what you said, for a given voltage the ring is maybe stable clocked a bit below core. I'm not turning up voltage just for the ring if I'm not doing it for the core, so I guess I'm stopping here. It's now passed 1h on Prime95 blend which is when it crashed at 4.2 ring. I'll give it another hour or so then put aida on it as a secondary test.

 

[Kenrou]

Out of curiosity, why only 1.25v ? You can safely go 1.42v on 24/7.

If it helps, mine was 4.7/4.1 1.32v, 4.7/4.7 1.33v Prime95 blend 2h stable.

 

[mackerel]

Power output... it is summer and hot. Let's say I hit 4.7 GHz with increased voltage, which is ball park of 10% more than 4.2 GHz. 10% is nice. But the increased voltage will mean power consumption will be far more than 10% higher. I wouldn't normally even be running it overclocked at all, but next week I'm taking part in a distributed computing challenge so thinking I might tolerate it for the short term to get a bit more of an edge but it has to be 100% Prime95 stable.

Also I suspect I had a bad (OC) sample. I can't even run stable at 4.5 with 1.35V set although I don't know what ring was set as I left it on auto... hmm... maybe I should look at it again.

 

[Kenrou]

Aie does look you lost the lottery as I said, stock for me is 4.1, and doesn't budge from there unless I pump it to 4.7ghz and above. Maybe set min/max manually to 4.0/4.1 ?

 

[mackerel]

Just tried 4.5 core, 4.0 ring at 1.35v like last time. Still p95 unstable.

- - - Updated - - -

Did you use LLC? I didn't, and also didn't monitor if there was much droop. Had enough for now, going to finish verifying 4.2.

 

[Kenrou]

My mobo has LLC to 5 (Auto, 1-8), gives me the steadiest settings, 1.32v BIOS reads out as 1.328v idle and 1.312 under load on HWiNFO64.

 

[bob4933]

Going from 1.25 to 1.42v is 39w (assuming a modest 4.7ghz). Going from 1.25 to 1.35 is less than 20w. Your concerns about heat seem to be a bit inflated.