Intel i7-4790K Core Speed vs Cache Speed

InfernoxCJC

Honorable
Sep 21, 2014
143
0
10,690
Hello again everyone!
I do apologies for the amount of threads I have been making recently, hopefully I will go away soon enough! =P

I have been having a fair amount of trouble trying to get my Intel i7-4790K stable at 4.4GHz Core and Cache but haven't been able to pass Prime95 (27.9) or AIDA64 tests without crashing and have pretty much given up at this stage with my latest attempt with 1.25V on both Core and Cache resulting in temperatures hitting 90C during Prime95.

I built this rig back in October and thought I had it stable at 4.4GHz with 1.18V Core and 1.2V Cache but after attempting to move my RAM from 1600MHz to 2133MHz all my testing has resulted in crashes even back at the original Overlock so clearly I messed something up back then.

The system is definitely stable at stock so I suppose my trouble is down to a combination of poor luck in the silicon lottery and the weak link in my system being its cooling (CoolerMaster Hyper 212 Evo) which, when I was first buying parts I thought to be a fairly good cooler.

Anyway, down to the point, I understand that lower Cache Speeds allow for lower Core Voltages and allow for "easier" overclocking. My question is, how much, if any, of a performance drop would I see in CPU intensive games if I aimed for 4.4GHz Core but left the Cache at stock (Auto) settings? And what if stock Cache allowed me to hit 4.5GHz Core for example?

Conor
------------------------------------
TL;DR
What would the performance difference between 4.4GHz Core/4.4GHz Cache, 4.4GHz Core/4.0GHz Cache and 4.5GHz Core/4.0GHz Cache be in gaming scenarios?
 
I have become a bit jaded on the subject of haswell cooling for overclocking.
How high you can OC is firstly determined by your luck in the bin lottery.
I had high expectations from the Devil's canyon parts and their better thermals.
I found out that the thermals really do not matter unless, perhaps, you are a competitive overclocker.
Haswell runs quite cool, that is, until you raise the voltage past 1.25v or so.
Once you go past 1.3v, then you really do need very good cooling to keep stress loads under say 85c.
But, the consensus is that voltages higher than 1.30 are not a good thing for 24/7 usage.
I have been unable to find any official Intel recommendation on what is a safe vcore limit.

My advice is to reset all to default/auto and start again.
Stress test with OCCT which will be more representative of real world instruction patterns.
OCCT will stop the test at 85c.
Gradually raise your multiplier until you reach Vcore of 1.25.
At that voltage, you should have no heat problem.
What multiplier you get at 1.25v will be determined by the quality of your particular chip.

I might target ram to run at 1.5v 1866. Past that requires higher ram voltage.
I found that higher speed ram would limit the max oc a bit. That tradeoff is not much worth it if it costs you a single multiplier.

I would not fool with cache or other more esoteric knobs.

If you are a competitive overclocker, anything goes.
But if you are just a user interested in good performance, you can very well run at 4.0/4.4 turbo which is stock.
 
1. I find P95 and AIDA 64 useless for modern CPUs. If you are interested in getting your name posted on website overclocking leader boards, I suppose it's productive but you can be P95 stable and fail in an application benchmark (i.e. Real Bench) so why bother. Both P95 and AIDA 64 will create CPU usage scenarios which no combination of real world applications / games can produce.... it's validity as test mechanism is therefore suspect as you will unnecessarily limit your OC by looking for stability and temps at "imaginary scenarios".

2. Did you lose a "1" somewhere.... I presume you mean 1.18 VCore not a chip frying 1.8. Any P95 over 26.6 includes AVX and if using adaptive voltage control, it will add about 0.13 volts whenever AVX is present. For this reason, mist folks use 26.6 but again, how do you call your AVX equipped CPU stable when you don't test with AVX.

3. First off, set VCCin to 1.9 or your OCs above 4.5 or so will go no where.

4. What is the rated XMP speed of your RAM ?

5. The CM Hyper 212 is a "great budget cooler".... it is not a "great cooler"

6. Cache does not have a big impact on gaming .... image editing, yes.

7. The default clock of the 4790k is 4.4 Ghz....so that is not an overclock. If just doing gaming, you can leave cache at default.

Drop AIDA / P95 and download Asus Real Bench as well as HWinFO64 for monitoring. Set all BIOS settings to default and save / exit. The:

1. bump the CPU multiplier from Auto (44) to 45
2. voltage to Adaptive (or whatever your Mobo vendor calls it)
3. VCCIn to 1.9
4. Nothing else

and run Real Bench Benchmark.... if ya pass, do the 2 hour test. If ya fail bump Core voltage by 0.025, nothing else.

We do not want RAM running at XMP until we have determined minimum core voltage for each multiplier. Once that's done you can dry dropping voltage by small increments..... for example if 1.225 didn't work and 1.250 didn't, then I'd try ha;f way between at 1.237 ... continue to narrow the gap down till you happy.

At that point you can switch to XMP profile starting at the XMP voltage and going up as necessary till ya RB stable.

Note: RB includes AVX and other modern instruction sets but uses them to an extent found in real applications. Voltages will spike 0.13 or so when AVX is present. At 46 multiplier on 4770k, my voltages will spike from 1.375 to 1.48 - 1.50 for a millisecond or 2 when AVX instructions are present. Temps however have yet to break mid 70s. The DC CPUs run much cooler than HW

Finally you can try to drop VCCin from 1.9 to 1.86 or so and if it fails "split the difference again (1.88) till ya find the lowest stable VCCin

Rinse and repeat with higher multipliers if ya temp levels stay south of 80C.
 
Thank you for your replies geofelt and Jack!


geo:
When you say "Gradually raise your multiplier until you reach Vcore of 1.25" do you mean leave the voltage on Auto and increase the Multiplier until I hit 1.25V or set my Core Voltage at 1.25 and keep increasing the Mulitplier until I crash? I ask this because at stock (according to HWMonitor) my minimum VID is 0.719V and my maximum is 1.281V


Jack:
I have 8GB of Kingston Hyper-X Beast rated at 2133MHz

You say that 4.4GHz is not an Overclock but from my experience it is because the stock value is the "Turbo" speed of the CPU which it does not always run at, I saw a decrease in the region of 1000 points for example when running 3DMark FireStrike using stock rather than 4.4GHz

So you think I should ignore the XMP Profile until I reach a stable CPU Overclock yes? Furthermore, you suggest, at first, I leave the Adaptive Voltage at whatever it defaults to? Finally, by XMP Voltage do you mean that of the RAM?


Thanks so much for your help, both of you! <3
Conor
 
Using the manual overclock where you machine runs always at 4.4 Ghz is not recommended. You want to use Intel speed step and scale down to 800 Mhz at idle. This is accomplished by using "adaptive" voltage control (an Asus term) .

Puzzled as to why you saw a significant change in a GFX benchmark with different CPU settings.

yes, leave XMP off until you settle on mutiplier and vCore....remember RAM is controlled on CPU by the IMC and extra RAM load increases CPU memory management load.

We did a lot of Asus Z87 builds and I gave this out to peeps we built for. I would do the 42 multiplier with them here and they coulkd go on at home "at their own risk" to their liking.

It's not intended to be a guide for those looking to squeeze every Mhz outta their CPU ... it's intended to get you "pretty close" w/o a lot of time and effort. It wouldn't be much different for DC 4790k as for HW 4770k except of course:

1. Waste of time to do 42 - 44 multipliers as turbo on the DC is already 44.

2. DC is a lot more heat efficient so the cooler recommendations can prolly be moved up one voltage level.

Of course if not an Asus board, you will have to adjust for your MoBo vendor's BIOS terms and settings. Copy / paste follows:

==============================================

This is my personal "Minimalists Guide to Haswell Overclocking on Asus Boards". I don't have the patience to invest 100 of hours but I admire those who do. Using this method, most can knock it off in a weekend.

1. Stop using AIDA, Prime 95 or anything else like that. Download RoG Real Bench, HWiNFO64 and if ya want something extra Intel ETU.

http://rog.asus.com/forum/showthread.php?43233-Realbench-v2-Discussion-Thread-Download-Links
http://www.hwinfo.com/download.php
http://www.intel.com/content/www/us/en/motherboards/desktop-motherboards/desktop-boards-software-extreme-tuning-utility.html


Usage of testing and monitoring programs:

When you open Real Bench, move both windows to left side of screen. Open HWiNFO64, run "sensors only", you may get a pop up asking whether to disable reading the Asus EC chip (or soemthing like it), click "Disable this sensor". Move the HWiNFO64 window to upper right hand corner of screen. Stretch bottom of window to full screen height. Make the following changes:

-Right Click on "System" right at the top, select hide.
-In the next section, hide the last 4 lines starting "Core CPU Thermal Throttling" (if you watch temps, this is useless)
-Skip over the next section and Hide the section after that (section includes CPU Package thru DRAM Power)
-Now the whole reasons we did that was so you could see everything you wanna see at same time. You should be able to see Vcore 0, 1 and 2 at -the bottom of the window. If not hide a few more lines. Save and Quit will save your edits.


2. I am going to assume that you want your PC to power down and reduce voltages when not needed so for this we'll use Adaptive settings. Adaptive will throw an extra 0.10 to 0.13 volts at your CPU when AVX instructions are present. Again, I would NOT use P95 or AIDA on this setting without constant attention.

3. After setting BIOS to defaults, Input the following settings and then right click on them to add them to your favorites page. This will allow you to access all the settings you need to without bouncing all over the BIOS:

AI Overclock Tuner = Auto
1-Core Ratio Limit = 42 (all others should automatically change with Sync all cores selected above)
Max. CPU Cache Ratio = Auto
Min. CPU Cache Ratio = Auto
Fully Manual Mode = Disabled
Core Voltage = Adaptive
Additional Turbo Mode CPU Core Voltage = 1.200
Core Cache Voltage = Adaptive
Additional Turbo Mode CPU Cache Voltage = Auto
Eventual CPU Input Voltage = 1.90
DRAM Voltage = Auto

I'd suggest taking a screen shot (F12) of the favorites page when ya have successfully passed the stress tests.

4. Open Real Bench, select Benchmark Tab Check only the last box. Open HWiNFO64, run "sensors only" as described above. Start Real Bench and don't touch mouse till finished. Observe voltages and temps. If you can get thru these 2 minutes, your close.

5. Then try checking all 4 boxes and run again NOTE: During the 3rd test Open CL will send AVX instructions to CPU; pay close attention to Vcores as they will spike as described above. If passes.....

6. Switch to the Stress Test Tab and select the amount of RAM you have in your system and 2 hours..... (Note: If you plan on raising cache and / or RAM after a run, I will usually save the two hours and skip this step until I have Multiplier / Cache and RAM speed at my targets.

7. If at any point you fail, up Core Voltage to 1.225 (+0.025)...Always watch temps and stop tests if you reach temperatures of concern (> 85C in my book). Record the following:

42/A/A/Auto - Shorthand for 42 Multiplier / Auto Max. Cache / Auto Min, Cache / Auto DRAM setting
Actual RAM Speed - i.e. 1600
Ambient = Room Temperature
Coolant Temp at Idle = Requires a sensor
Idle Core Temps Before Test on Each Core = i.e. 25, 26, 24, 22
Average Core Temps for Each Core During Test = i.e. 59.6, 58.2, 52.7, 49.4
Max Core Temps During Test on Each Core = i.e. 65, 62, 59, 54
Settings you input in BIOS for VCore, VCC Ring (Cache), VCCIN (Eventual), DRAM i.e. 1.2000, Auto, 1.900, Auto
Actual Readings in BIOS for VCore, VCC Ring, VCCIN, DRAM i.e. 1.040, 1.122, NA, 1.671
Actual Readings in HWiNFO64 for VCore, VCC Ring, VCCIN, DRAM i.e. 1.296, NA, 1.920, 1.681
Highest Voltage Reading on any Core During each of the 4 Benchmarks, i.e. Image Ed. 1.200 / Encoding 1.216 / Open CL 1.296 / Multitask 1.248

7. If at any point you fail, up Core Voltage to 1.250 (+0.025). If ya fail again, go another notch (1.275) but I'd stop there.

8. Once you pass, it's time to consider cache voltage. Some are content to leave at Auto (39) as it affects very, very few applications (skip to step 9 if this is you), others try and get as close as they can to the CPU Multiplier. If you want cache up, go to 42/42/42/Auto. If ya fail, bring up cache voltage in same 0.025 increments.

Settings will look like this when starting:

Max. CPU Cache Ratio = 42
Min. CPU Cache Ratio = 42
Additional Turbo Mode CPU Cache Voltage = 1.200

9. Once stable, it's now time to get ya RAM up to its rated 2133, 2400 or whatever. Change 1st setting above to XMP

AI Overclock Tuner = XMP

Referring back to step 6, this is the point I normally do the 2 hour test when I am "done" with a given multiplier. So run the 2 hour test here, followed by an 8 hour test w/ Intel ETU.

10. If ya fail.... up ya voltages as per above..... as long as things don't get too hot.....see limits below. If ya pass, it's time to see if we can lower temps and voltages. I dunno if it matters what order ya do it in but I did VCCIN 1st till I failed then bumped up till I got lowest stable setting. Then did VCCring (Cache Voltage Setting in BIOS) till I got lowest stable setting....and finally VID (BIOS CPU Voltage setting) last. I leaped in "half" amounts.

For example.... Default VCCIn is reportedly less than 1.8 .... so if 1.9 worked, i went "half way" to 1.85 .... if 1.85 failed, I went halfway between known good and bad to 1.875 ....same deal with VID and VCCring.

11. With the 42 series if tests complete, "rinse and repeat" with steps 3 thru 10 after moving up to CPU Multiplier to 43, then 44 or as high as you are willing to go. At 46 multiplier I found 1.9 VCCIN to be inadequate.... this is the one voltage I found that going too high or too low is problematic (other than heat and maximum upset voltage limits of course). I went to 1.98 (last yellow setting) and it was too low..... 2.08 was too high. 2.04 worked for me w/ 46 multiplier,

12. These are my settings to give ya an idea of luck I had .... your mileage will vary. Asterisked ones are those I didn't go back and try and get better temps / voltages.

42/42/42/XMP (2400) *
VCore 1.200
VCC Ring 1.200
VCCIN (Ev) 1.880
DRAM 1.700
Avg Max Core Temp = 59.0C (28.0C Ambient)

43/43/43/XMP (2400) *
VCore 1.225
VCC Ring 1.225
VCCIN (Ev) 1.880
DRAM 1.700
Avg Max Core Temp = 56.5C (27.4C Ambient)

44/44/44/XMP (2400) *
VCore 1.260
VCC Ring 1.260
VCCIN (Ev) 1.880
DRAM 1.700
Avg Max Core Temp = 59.0C (24.5C Ambient)

45/45/45/XMP (2400)
VCore 1.325
VCC Ring 1.325
VCCIN (Ev) 1.880
DRAM 1.700 *
Avg Max Core Temp = 63.0C (22.9C Ambient)


46/43/43/XMP (2400)
VCore 1.385
VCC Ring 1.385
VCCIN (Ev) 2.020
DRAM 1.700
Avg Max Core Temp = 69.0C (23.0C Ambient)

46/46/46/XMP (2400)
VCore 1.385
VCC Ring 1.410
VCCIN (Ev) 2.040
DRAM 1.70
Avg Max Core Temp = 72.0C (24.4C Ambient)


13. As for cooling / heat / voltage concerns

Here's Asus recommendations:

A very good air cooler is required for voltage levels above 1.15V.
1.20V-1.23V requires use of closed loop water coolers.
At 1.24V-1.275V dual or triple radiator water cooling solutions are advised.

My thinking is:

Up to 1.200v = Very Good Air Cooler (Hyper 212)
Up to 1.250v = Best Air Coolers (Phanteks PH-TC14-PE, Silver Arrow or Noctua DH14) ....... Dual 140mm CLC / AIO Cooler w/ 1500 rpm fans (Corsair H110)
Up to 1.275v = Extreme Speed Dual Fan CLC / AIO w/ 2700 rpm fans (too noisy for most folks)
Up to 1.287v = Best air coolers (Cryorig R1 / Noctua DH-15)
Up to 1.300v = Swifteh H240-X
Up to 1.325v = Custom Loop w/ 15C Delta T (3 x 120mm / 140mm) *
Up to 1.400 = Custom Loop w/ 10C Delta T (5 x 140mm or 6 x 120mm) *

* At this level having the GPU(s) also under water is assumed

Also, if you are not running AVX, you can add as much as 0.10 to all those voltages.

14. NEVER WALK AWAY from your machine while stress testing until you are sure that temps have stabilized.
Be AWARE if test uses multiple instruction sets like Real Bench who throws out its hardest load voltage wise with the 3rd test in the Benchmark but the 4th test results in higher temps.

Remember some AVX instructions are present during RoG Real Bench type loads which will raise VCores by 0.10 to 0.13 for short periods.
I would not suggest running Prime 95 w/ AVX under adaptive under above conditions.

15. Having 4 sticks of memory will hinder ya OCs a bit.

16. If ya want the best OCs ya machine can get, this is not the guide to use. If ya wanna get it done over the weekend in between taking work home, course work, Honey-Do Lists, Daddy Taxi and other life demands, this may get it done in a weekend :) .

WORD OF WARNING: Some of us are having problems with the BIOS clock freezing and a suspected cause is the use the saving, loading and backing up of OC profiles in the Tools section of the BIOS. I would avoid use of that feature until such time as the cause is confirmed or a fix is available.



 
Thanks a bunch for that detailed guide Jack, will go through it tomorrow!

Just to clarify, I am still using Intel SpeedStep, I simply meant I was using a maximum Core Speed of 4.4GHz rather than a Turbo Core Speed
Using the stock Turbo settings, when running a Stress Test such as P95 the Core would run at 4.0GHz only reaching 4.4GHz under less intensive scenarios and FireStrike does have a dedicated CPU Test which resulted in the difference.

I am using an ASUS Z97-A so my settings are near enough identical to those you used, I will do the tests tomorrow and get back to you in the evening, thanks so much for your help!
Conor
 
Hello again Jack!
I have been attempting to run the RealBench two hour Stress Test having already passed two hours on OCCT and have had two Display Driver crashes (most recent Nvidia drivers)

Can these crashes be caused by an unstable CPU Overlock or are they solely GPU/Driver related?
 
Thanks again for your help Jack!

Thus far I have been struggling to find stability with higher Cache Speeds but as you say and with what I said in my original post it is arguable that this isn't really that much of a problem.

I mostly use my system for a wide variety of games types along with some CAD and programming software while I am at University, how much, if at all, will stock Cache speeds hinder my performance?

Would it be better to aim for say 4.5/4.0 (Core/Cache) or maybe 4.4/4.4 and even 43/43 if reaching 4.4 is not possible?
 
Can't believe noone has pointed this out yet, but 1.2v is the upper limit of cache voltage. It's actually too high for 4.4ghz, let alone 24/7 operation. Keep it down to 1.05v, 1.15v at most. This will not only reduce temperatures but likely also enhance stability. Even if you may need to turn cache speed down a bit, what's the deal? I doubt that many people use software that requires a high system ring speed.
 
Is that a fact? I've never seen anything about that before, I mean prior to this thread I have been running my i7-4790K at 4.4/4.4 at 1.80V/1.23V (Core/Cache) and didn't have any issues that I am aware of...

Speaking of this however, if I set a (Manual) voltage of 1.2V Cache in my ASUS Z97-A, HWMonitor shows an actual voltage of 1.23V, this +0.03V increase is present no matter what (Manual) voltage I set for Cache...
 


Yeah, 1.2v is the maximum recommended for testing purposes, 1.15v recommended maximum for safe overclocks. 1.2v cache voltage is around 1.45v vcore on dc cpu's, some do actually work nicely at around that much but a lot stop working or even die from it. Probably the same as with input voltages, some i7 4790k scale greatly with lower vccin, some need their 2.0v to get stable on high overclocks.
 
Forgive me, isn't VCore a totally separate setting from Cache Voltage?

Thank you for your input Dubble, and I am by no means questioning you as I will happily admit that, while I feel I have a better understanding than most, I am still miles from the level of understanding most of you avid solution finders here at Tom's are but I am eager to learn!

It's just that, I have never seen any mention of this lower "safe" limit for Cache. There seems to be a rather unanimous agreement that 1.3V Core is the limit for safe, everyday usage (assuming you have the appropriate cooling solution) but I have never seen mention of this 1.2V limit on Cache, on top of which most, if not all the examples of i7-4790K overclocks that have included Cache, such as those posted by Jack previously, have used voltages exceeding this value.

I am not attempting to undermined you, I am simply trying to maximise the performance of my system while ensuring its longevity and improving my own knowledge.


Thanks again,
Conor
 


Yup. The maximum safe vcore is generally considered 1.45v under air/water. For the vring (cache) the safe maximum is considered to be 1.2v, no matter your cooling. Above is for testing, not for everyday use. Electromigation may "kill" (or well, degrade) your chip quite fast at those voltage levels, no matter the cooling. For safe everyday use, most people prefer to have them under 1.3v/1.35v vcore and 1.15v/1.2v vring.

I didn't really look for links on vring recommendations (due to those being rather unpopular and being short on time), but had one bookmarked in my browser: http://extreme.pcgameshardware.de/overclocking-prozessoren/277031-howto-intel-haswell-oc-guide-inkl-haswell-cpu-oc-liste.html
In case you are questioning the OP of that article, check up on overclocking world records and you'll see his name fairly quick. It's german, but google chrome should get that understandable.

If you want to just get your system stable and quick for everyday use, not using very specialised extremely cache intensive software specifically catered towards haswell cpus, you probably won't notice a difference between. 3ghz and 4.4ghz cache speed. You should get around 4.2ghz on 1.05v vring though, so there's no point going lower.

One thing that's definitely worth a try is messing (lowering) your vccin, input voltage. Some devils canyon cpus seem to have a voltage regulator different to haswells that scales greatly to lower input voltage, I personally saw one of my i7 4790k's requiring only 1.21v vcore for 4.7ghz applying 1.45v(!!!) vccin instead of 1.33v it would require with the standard 1.8v. Still wouldn't get to 4.8ghz stable, though.
On the other hand the other two I tested didn't scale very well to it. I'm personally running mine at 0.98v vcore/1.4v vccin (4.0ghz) with 1.0v vring (4.0ghz) now, which has definitely helped my temperatures (cooler doesn't go over half RPM, which is great because I hate noise). But it does require at least 1.96v vccin to get "stable" (applications) at 5ghz with 1.50v vcore. Or 1.88v (possibly lower, didn't test) vccin for 4.9ghz with 1.4v vcore. Lowering the vccin on this cpu at high clock speeds only got me crashes.

Also, no worries on undermining me (lol), I'm not exactly an authority for overclocking or anything. Just happen to be into the topic as a hobby as well as having a lot to do with computers at work.