[RedKnight7]

Here's some stats out of Darkwizzie's spreadsheet, from his Haswell Overclocking Guide With Statistics. Help out - post your junk there.

The inset's also attached as a pic so you can see it better. So is the spreadsheet, if you want to play with it. But it's static... Darkwizzie is always adding to the original with more. You too can be a proud datapoint.

Here tis:


DarkwizziesCPUstats4-17-2014.jpg 394k .jpg file


DarkwizziesOverclockingResults.xls 88k .xls file

• On average, overclocking takes +0.023 VID per multiplier (e.g. 44 to 45)
• This is true for 4670K and 4770K. In fact, they only differ by 3% (0.0227 vs 0.0234).
• The two chips overclock quite similarly, with a slight nod to the 4770K (average overclock for 4770K is 4.566 versus 4.526 for 4670K). But this is only a 0.04 Ghz difference even though the chips have a nominal 0.1 difference (3.4 Ghz vs 3.5 Ghz).
• You might consider a 1.3 VID not so safe. If so, 54% of folks (88/163) played it safe, with VIDs below 1.3.
• My chip sucks! (red squares)

Of course, this is not a comprehensive systematic review. Still, it's what Haswell clockheads here have found.

Anybody wanna see more? Core-to-uncore ratios and voltages, VID to VCCIN, etc.

 

[Cyro999]

VID to VCCIN would be interesting! Thanks for this

 

[RedKnight7]

Here you go Cyro,

Here are some more stats from Darkwizzie's spreadsheet in his Guide ...

I examined VCCIN and Uncore stats. Since no one else has shown interest, I won't go into length here. For extended details, see the graphs and tables in the attached spreadsheet (note the three Tabs).

4670K and 4770Ks were lumped together, since the OP found little difference voltage-wise:

• VCCIN (also known as Vrin or ASUS Eventual Voltage) seems constrained by a 2.0 Volt barrier. (My BIOS shows values lower than 1.99 as yellow, but goes violet for 2.) Only 6 out of the 89 people who included VCCIN were over 2 volts. I looked for correlations with VID or Core Multiplier, but there were little or none. Between the 2-volt mental barrier for some, and others who probably didn't spend much time on VCCIN, I don't think there's much advice one can glean from this data. See the spreadsheet for supporting info.
• Most VCCIN values were between 1.8 and 2.0 volts, with an average of 1.89. As an offset from VID, most were 0.5 to 0.7 more than VID.
• VCCIN did not scale proportionately with VID (as a fixed percent but higher absolute value). In fact, the opposite was seen. Undoubtedly due to the 2-volt mental barrier.
• VCCIN also did not correlate well with Core Multiplier (R^2 0.22). As you might suspect given the highly variable OC quality of Haswells, VID vs Vrin is a better predictor here (see first bullet), but not by much (R^2 0.36).
• Uncore (a.k.a. Cache) Multipliers did not correlate well with Core Multipliers. 158 of the records had Uncore data.
• Only 8 (5%) had Uncore going as fast as Core. It's a surprisingly small percent, to me. Of course, no one had Uncore going faster than Core.
• Otherwise, there is practically zero correlation between Uncore and Core Mults (R^2 0.06). A chart of the one versus the other looks like a shotgun blast. Maybe some folks who drove their Core high compensated by driving Uncore low? This would really mess with correlations.
• While the slowest Core Multiplier for the 158 records was 42, the slowest Uncore was 30. There are a LOT of slow Uncores; a fifth (31/158) were below 38, which is the Auto default on my motherboard(!). The average Uncore mult was 40.1 and median 40, versus average Core mult of 45.5 and median 45 for the same records.
• On a pairwise basis (for each CPU), Uncore was an average of 5.4 Mults less than Core, Median 5. The max difference was 17 for someone with Core 50 and Uncore 33. So help me God.
• Oddly, the few folks who matched Uncore to Core only did so in the range of 44 to 47, whereas the entire Core Mult range is 42 to 50 for these 158 records. Again it makes me wonder if we aren't seeing hardcore folks pushing Core Mults high but constraining Uncore Mults to compensate on the one hand, and conservative folks actually (mistakenly?) pushing Uncore Mults very low "to be sure" on the low end. Taken together, it means there's practically no correlations to be found in this data. (See graphs in spreadsheet.)
• There were 126 records with Uncore Mult and Uncore Voltage (not counting 30 with V_uncore set to Auto). The average slope is 0.0124 Uncore volts per Uncore mult, but the chart suggests a poor correlation (R^2 0.33). I have to wonder if it's obfuscated by factors mentioned previously.

I'd like to repeat that I'm new to overclocking, so if any vets want to step in with observations or corrections, please do.

Also, let me thank Darkwizzie and the other Guide contributors once more. Without them, this would not be possible. Please post your Haswell OC results in that thread per the instructions there - share and share alike.

Good multipliers,

RedKnight7

DarkwizziesOverclockingResultsStats4-30-2014.xlsx 105k .xlsx file

 

[Cyro999]

Quote:

Again it makes me wonder if we aren't seeing hardcore folks pushing Core Mults high but constraining Uncore Mults to compensate on the one hand, and conservative folks actually (mistakenly?) pushing Uncore Mults very low "to be sure" on the low end.

You can't really clock core higher because you're running a low uncore. The guys running ~33-40 uncore are just doing it because they didn't bother to put it back up because it's essentially irrelevant for almost every task
Quote:

The average slope is 0.0124 Uncore volts per Uncore mult, but the chart suggests a poor correlation (R^2 0.33). I have to wonder if it's obfuscated by factors mentioned previously.

It is, many people using [email protected] as the guide says to start with. It's overkill by intention

Thanks for post!

 

[mav451]

0.02v per mutli increase seems way too small.
To be fair though, the pool of OC results is not standardized by any single metric, but my personal observation is that the increase is closer to 0.04 or 0.05v.

0.07v per increase is definitely out of the norm though, sorry man haha

 

[Cyro999]

Quote:

0.07v per increase is definitely out of the norm though, sorry man haha

It happens quite a lot actually, you just have to be closer to chip limits to see it - like further than you have pushed yours. If you have a multi close to 1.3vcore and you're trying to increase over it and expect less than ~0.05-0.07ish, you're probably gonna have a bad time

 

[RedKnight7]

Cyro,

Excellent point - that the Guide recommends keeping Cache low. I had not taken it so literally, and assumed most folks would still try OCing their Uncore at least a little. But the charts entirely support that a fair number of folks not only didn't try, they did the opposite and simply dropped or Auto'd the Uncore mult. All in all, it dashed any correlation.

But now that you mention it, look how in fact it's strong supporting evidence for Darkwizzie's finding and recommend that cache frequency has little impact. (Not zero impact, but not much.) Surely most everyone setting their Uncore low did some testing to make sure it didn't hurt them (much). And plenty then stayed with a middling or even deliberately low setting. Strong evidence indeed, I should think. Nobody here is a dummy.

Hi Mav,

Well, the 0.023 VID per Core Mult is a regression line through everyone's data. Good chips, bad chips, conservative folks, madmen. Like you say, it's not standardized by any metric. It's also not standardized by goal or other things.

And each person only submitted one point. So really we're defining a "comfort surface", not individual CPU curves. It's fine for seeing what one settled on versus what others did, but it's not so good for choosing a VID-vs-Mult line, per se. Like you said.

And true - my chip really sucks. If I tried going much faster, it got over 100 C. I chose to play it safer, but it's still high by anyone's standard (1.431 VID for 44 Mult; 80 C HWiNFO CPU Package average on x264 v2 test). I can afford another chip; in fact, I almost relish the chance to try a better one. But most folks are more cautious. Have you pushed yours to 100 C under stress?

Ok then

 

[error-id10t]

Just few things.

ASUS VCCIN = CPU Input Voltage (ROG boards may have initial and eventual). I also don't expect you to find any rhyme or reason for a particular value to be honest. 1) because the range available is large 2) there appears to be no penalty for temps. I do have a theory which I can't backup but have noticed twice myself; with too low of a VCCIN which is still stable not to cause any BSOD, your performance is hindered compared to a higher value. It's a small penalty. There is a cut-off where this stops though so simply maxing it has no sense.

As already mentioned, some tests like cache while others don't. Some tests are faster with 45/43 vs. 46/35. I don't believe most people test "performance", they pass x264 and then do their daily tasks. I also don't believe most people come back to update if they had problems weeks/month later. Again, nothing can be done as that's what was recorded with no further updates. Further, the earlier x264 was quite useless to measure stability (the default x264) but the later modded version(s) are better - there's no mention who used which.

Lastly, I also second that the "standard" VID difference per core - once you reach higher multi's - will be closer to 0.05v. Of course you can only go by what has been charted. For me as an example the jump from x45 to x46 has been 0.06v (1.35v to 1.41v).

 

[RedKnight7]

Thanks Error, I think all that you write sounds good. Please let us know if you find out more about your theory, or effects of temps (versus voltage).

Hey Mav, in re-reading what I posted, I think I wrote poorly... when I wrote "Ok then", I just meant "Ok, let me stop here". I didn't mean it in the sense of "you didn't try that, now did you?"
Would hate to offend

Thanks for posting

 

[BoredErica]

I randomly Googled the OCN version of my name and I found this thread.

Hiya!

 

[Cyro999]

Hai!

 

[RedKnight7]

Heya!

Thanks again for your spreadsheet and Guide!

 

[BoredErica]

Cursed OCN won't let me add an underscore to my name.

Shoulda thought about that when I first chose to use the name.

And you're welcome again!