Core i7 920 Overclock 2.66GHz -> 4.22GHz Stable

[Kougar]

I guess I'm safe to brag every now and then. Rob asked me if this overclock was stable... well, lets see. There's actually a pretty lengthy story for this build, it took much longer than it should have before I could even get it started, but I won't get into that here. I assume pics will be wanted... I'll have to take some, sorry!

After spending a full day gutting, tearing down, cleaning out, and eventually rebuilding not just this system but my trusty Q6600 platform into a HAF 932 case (my now backup system) I was quite eager to sit down and have some fun, not to mention have a computer to use again! The case and a few parts were resused, the entire watercooling loop stripped down, cleaned, and rebuilt with some improvements and new tubing. I did run into some problems mounting the Apogee GTZ waterblock, if I had known in advance I would have gone with the D-Tek Fuzion. My build would have died on the spot if I hadn't also just happened to ahve bought a Thermalright LGA1366 mounting kit, which fixed the three main problems I ran into with the Apogee GTZ. Again, I simply recommend users go the D-Tek route on this one as I am less than impressed with the GTZ's mounting design. If I could've gotten my money back I would have waited the required time for the D-Tek unit to arrive before even completing the build, it was just that bad of a mounting design flaw.

To avoid the Intel RAID nightmares I used just a single 750GB Seagate drive. I'm done with RAID, and in the future the next system upgrade will be to an SSD (Intel is releasing 34nm flash SSD's around the time Windows 7 will go gold, very convenient)... the complete ease of use and lack of problems only tells me I made the right choice forgoing RAID this time around.

I built this rig just after Windows 7 Beta went live, so after firing up the system and installing Windows 7 Beta 64bit, I manually set the correct stock RAM timings in the BIOS, rebooted to Windows, and gave IntelBurn a quick run:

Here's IntelBurn.

Notice anything odd? Hmm, wait a minute... I haven't even overclocked yet but the system isn't stable!? At this point I got a sinking feeling about my untested computer build, that dreaded feeling that almost anything could be wrong and I might never find the cause. Oh, did I mention this EX58-UD5 motherboard was a $200 Open Box buy, and way past the return date?

Not to keep in suspense this one I'll chock up to a bit of luck and a keen eye... I remembered distinctly that Gigabyte BIOS's only allow vdimm settings of 1.64v or 1.66v... not the 1.65v called for by the RAM. So playing it safe I had used 1.64v, I mean how could 0.01v be the absolute difference between stability and instability with 6GB of RAM? On a hunch I changed the setting to 1.66v it, and suddenly there were no more errors. Hmm, lesson learned...


Now with that out of the way and a backup system to use it didn't take me long to lose my good manners and decide to overclock the new system before it'd even been breathing for a full day. Not in any particular order, most of my tests were conducted with LinX as IntelBurn is only able to test four of the eight threads. These are my more recent screenshots, as I think all of them were Windows 7 RC:

24 Hours Prime stable, check



20 4GB runs of LinX under the new RC build of Windows 7, check



Six loops of 3DMark Vantage with all settings maxed out, check



Several lengthy gaming sessions with good friends on Left 4 Dead with settings all maxed out lasting a few hours, double-check

(Sorry, no screenshots )

Two full loops of Memtest86+ 2.11v, check

(No photographs either)

Oh, I did say 5GB problem sizes of LinX are much more effective for finding instability, so here's a >5GB problem size of Linx. Each run took over three minutes, 20 runs over an hour. I have done a final 25 run test using a slightly higher problem size that required almost two hours, but I didn't screenshot that for some reason. Because paging the entire 6GB of RAM will make the system laggy a sin I haven't bothered to do it again for a proper screen, but I probably should since I've tightened the RAM timings a bit further yet again...



After reading about how Windows 7 changed the system performance scoring, I had to give it a whirl:



5.9 is the highest a mechanical hard drive can score as I remember, only SSD's have the ability to score to 7.9 depending on how well they do in stuttering and other write/read tests. Believe it or not a GTX 260 didn't even get a 7, and a 4.2GHz Core i7 920 could not even get a 7.9... simply wow! I wonder what kind of system requires a 7.9 processor score... dual-socket Nehalem? After I tuned my RAM timings further, my RAM score finally did reach the maximum score possible of 7.9 shown in the screenshot.

And for those with a keen eye, or those still reading this far, here's another juicy tidbit for Core i7 overclockers they'll need to know. I could do all of the above using a Vcore of 1.362v, with LoadLine Calibration Enabled. Obviously this removes the vdroop protection Intel uses. With Loadline Calibration Disabled, I could do all of the above using a Vcore of 1.42v. My screenshots may show either voltage as I haven't bothered to sort them and I haven't decided on if I prefer LLC on of LLC off. I've run the tests on both settings. I know LLC off is easier for the PWM and power circuitry in general, and safer for the CPU, yet it's a "higher" vcore (or so one might first think).

Temps remain the same at these voltages. By comparison, 1.38v with LLC Enabled was 3C hotter on all cores than 1.425v with LLC off during LinX loads So there is no real temperature advantage gained for using LLC.

Respectively, a Vcore of 1.35v and 1.40v tests as unstable. VTT is 1.30v, IOH is 1.20v. Again, respectively a VTT of 1.26v or a IOH of 1.16v would test as unstable, so this is as low as all voltage settings will go. I didn't need to play with the PCIe frequency either, for a mere 201Bclock that is simply not needed. For those curious, my CPU is batch: 3844a759

And Rob, Gigabyte UD4/UD5 boards will hit 220Block just fine at the right settings with the right BIOS. The board design limits Gigabyte to somewhere between 222 and 227 Bclocks, it's an almost universal limit due to the motherboard layout.

By all accounts my 4.2GHz overclock was one of the worst on XS for a D0 stepping Core i7 920... but I don't care as it's stable with everything I throw at it. Stable with HyperThreading and Hardware Virtualization (For my VMware Ubuntu 64bit install) both fully enabled, since XS users seem to prefer higher clocks without HT. There is quite a bit more good info I learned during my first forays into Nehalem overclocking, but I won't spill it all just yet...

 

[Rob Williams]

So playing it safe I had used 1.64v, I mean how could 0.01v be the absolute difference between stability and instability with 6GB of RAM? On a hunch I changed the setting to 1.66v it, and suddenly there were no more errors. Hmm, lesson learned...

Something tells me that the real voltage difference is larger than that. It would be very unlikely that 0.02v would cause the CPU to go from unstable to stable... just doesn't make sense. What I find a little bit odd is that on our RIIE motherboard, even if 1.65v is selected, the real value is much higher (like 1.75v).

Thanks a ton for this report though... incredibly useful information here (especially regarding Load-Line Calibration). Only thing I have to ask is... temps aren't an issue with an overclock like that at all? When I overclocked to 4.2GHz, it would hit the threshold fast (Thermalright Ultra-120), and though it wouldn't crash, it'd throttle like there was no tomorrow (tracking with CPU-Z). All of your screenshots show the benchmarks stopped, so I can't tell what kind of load temps you're hitting.

As for GBT's BCLKs... good to know.

When all said and done, 4.2GHz on that CPU is simply incredible. What a freaking value that is...

Edit: Could you please use smaller screenshots or thumbnails in the future? The forum is broken with such large images :-( I wish I could figure out how to have the forum auto-resize these things...

 

[Kougar]

Something tells me that the real voltage difference is larger than that. It would be very unlikely that 0.02v would cause the CPU to go from unstable to stable... just doesn't make sense. What I find a little bit odd is that on our RIIE motherboard, even if 1.65v is selected, the real value is much higher (like 1.75v).

I would not be at all surprised, I still would be skeptical myself. I just know it made all the difference in the world when using this RAM at stock 1600MHz CAS 7 settings. I wish this motherboard has measuring points like some other X58 boards, but I should be able to find the correct points to measure if I poke around the XS forums. I am curious to know as well...

Edit: Measured a few things, the only voltage chip I could find that was close was exactly 1.703v under load. I currently am testing something so my BIOS vDIMM setting is 1.68v, so I presume this was it. Next time I set the memory back to 1.66 I'll remeasure it. Because the CPU voltage is so high I'm still well within the 0.50v differential between the Vcore and Vdimm voltages.

Thanks a ton for this report though... incredibly useful information here (especially regarding Load-Line Calibration). Only thing I have to ask is... temps aren't an issue with an overclock like that at all? When I overclocked to 4.2GHz, it would hit the threshold fast (Thermalright Ultra-120), and though it wouldn't crash, it'd throttle like there was no tomorrow (tracking with CPU-Z). All of your screenshots show the benchmarks stopped, so I can't tell what kind of load temps you're hitting.

Realtemp does show you what the "peak" temperature reading was for each core in those screenshots. It will always record the absolute min/max values as long as it's left running. Ya can see it peaking at 80c for the hottest core and 73c for the cooler running core.

What's the temp threshold for Core i7 again, 90c on the cores? And how did you monitor the throttling, by using RMClock or something else? I don't like the temps reaching 79-80c, but only LinX or other Linpack based programs are capable of doing this so I can live with it.

Even now the temps range from 63-76 across the cores with my current Folding@home load... the low reading is for the one core that seems to run unusually cool that is also the only one not under a full 2-thread load.

When all said and done, 4.2GHz on that CPU is simply incredible. What a freaking value that is...

It did make the slightly higher cost of buying a D0 worth it, but frankly $266-300 for a Core i7 920 is still expensive in my opinion! I'm going to be watching Lynnfield very closely to see how those OC.

Edit: Could you please use smaller screenshots or thumbnails in the future? The forum is broken with such large images :-( I wish I could figure out how to have the forum auto-resize these things...

Yeah, I know the forum has the ability to auto-resize images. I had to edit all the screenshots because when I first made the post I assumed the forum would auto-resize them.

 

[Kougar]

As I said above RealTemp records the maximum/peak temperature so in effect it will show what the load temps were for the system. Still, since Rob asked here's a shot with Prime running:



Here's one showing LinX with all of my current (24/7 daily use) settings, 201Bclock for 4.22GHz



After doing a good bit more reading I've couldn't resist some more testing, and now am pretty sure I can get a 24/7 stable 205Bclock for 4.3GHz, possibly 210 for 4.4GHz if I pushed the vcore.



I got a little impatient and skipped 210Bclk and went for 215Bclock for the magic 4.5GHz, but alas it wasn't to be. To get 4.5GHz stable this particular CPU would need >1.5 vCore which I'm not willing to do. I'm not worried about killing the chip as that's unlikely, but degredation would be possible, and probable.



As far as temps go, if I don't undervolt the radiator fans temps stay in the low 70's even under LinX. With the fans undervolted to around 7v LinX is usually in the upper 70's, most other loads like Prime are around 70 or lower.

I did some stress testing where I overvolted and overclocked the GPU, then ran Furmark to dump heat into the loop, and then also ran LinX on the CPU... temps peaked at 83c. There was no CPU throttling whatsoever and no Turbo throttling either. This is what the i7 Turbo V8 program is for, from the same guy that created RealTemp. ASUS's P6T board apparently has some major issues with throttling Turbo under loads when the temps get warm, makes me glad I stuck with Gigabyte!

 

[Merlin]

I haven't tired OC'ing my i7920 as yet, on the MSI platnum board, it's a cheap board that I thought I would get just to get things going. But, it seems to be a good board after all and I'll just keep it.
So, now I'll look into some over clocking.
What's the main thing on these i7 chips to OC ?

 

[Kougar]

Your board will do a 200 Bclock just fine, but I haven't checked how MSI handles the Turbo mode. If you can you'll want to force it on to change your 20 multiplier to 21 for a free 200MHz.

As for overclocking Core i7's in general... it's much more complicated than it used to be, since there's every setting under the sun now available to you in the BIOS. Also you need to keep track of Uncore frequencies, QPI frequency, and so on...

Basically, CPU voltage, QPI Voltage (VTT), and a little IOH Core voltage should get you to a 200 Base Clock. I don't mess with the PCIe voltage, PCIe frequency, and the CPU PLL and QPI PLL voltage settings shouldn't need to be changed unless clocking above 200Bclks.

Edit: All of this assumes you have a D0 stepping processor, otherwise it gets trickier and the VTT voltage will need to be set much higher...

 

[Rob Williams]

And how did you monitor the throttling, by using RMClock or something else?

it'd throttle like there was no tomorrow (tracking with CPU-Z)

The "Core Speed" box is actively monitored and updates every 1s or so. During stress-testing, it would drop down 600MHz or more, so it was fairly obvious that temperatures were getting in my way.

It did make the slightly higher cost of buying a D0 worth it, but frankly $266-300 for a Core i7 920 is still expensive in my opinion!

I don't think I can agree on the expense issue... I truly find that to be a great value. That's an insane amount of power, right there. Really depends on what you're doing though, of course.

Yeah, I know the forum has the ability to auto-resize images. I had to edit all the screenshots because when I first made the post I assumed the forum would auto-resize them.

I looked all over and couldn't find that option in the admin section at all. vBulletin is extremely complex at times... so difficult to find a seemingly simple option.

I'm not worried about killing the chip as that's unlikely, but degredation would be possible, and probable.

I'd be the same way... and if using Prime gets your temps up to 70C+, I really don't think you want to be pushing it that much harder. TJ-MAX is 100C on Core i7, I believe, and according to temperature monitors, I've hit it, more than once (Everest 5.x). I'm desperate for better cooling. Anyone hear about Corsair's latest water-cooling unit yet?

 

[Kougar]

The "Core Speed" box is actively monitored and updates every 1s or so. During stress-testing, it would drop down 600MHz or more, so it was fairly obvious that temperatures were getting in my way.

Hmm, interesting. What temps was this occuring at? I found a nifty tool (Guy that wrote RealTemp) that monitors Turbo mode throttling, and Real Temp does monitor thermal throttling. Neither of them occur on my CPU and I've pushed it as high as 83c across 3 out of 4 cores.

ASUS's implementation of Turbo Mode isn't that great, it will rapidly drop/raise/drop the CPU multiplier between 20 and 21 once the CPU reaches a certain temp, without CPU-Z being aware of this occuring. Gigabyte, like ASUS, allows a BIOS option to force Turbo Mode on, however it will stay on up until the CPU itself begins to throttle, at which point it supposedly disables itself. If you didn't have CPUZ showing the loss of clockspeed, then the performance differences you spoke of regarding the throttling could have been their implementation of Turbo switching on/off when it shouldn't. Turbo mode does complicate CPU testing...

I looked all over and couldn't find that option in the admin section at all. vBulletin is extremely complex at times... so difficult to find a seemingly simple option.

I so know it's in there or available as an option, because some other vBulletin forums resized images for me. I just now made sure it was vBulletin and not phpbb.

I'd be the same way... and if using Prime gets your temps up to 70C+, I really don't think you want to be pushing it that much harder. TJ-MAX is 100C on Core i7, I believe, and according to temperature monitors, I've hit it, more than once (Everest 5.x). I'm desperate for better cooling.

Well, yes just that keep in mind the radiator fans were still around 7V for silent operation. I tried LinX @ 4.3GHz using 1.45v and temps were peaking at 81-82c with the fans turned up a little... and no throttling was occuring. The 100c thermal limit is not the Core temp, but the temp for the geometric center right above the IHS surface, so it really muddles things.

According to the guy behind Real Temp some users reported throttling at 80c, some don't report throttling until 85c or even higher. In my case there's zero throttling, I have never seen this CPU throttle itself to date during my overclocking experiments (again 83C is as hot as I've been able to force it by setting the fans to ~5V with a full GPU load). And to answer your question I don't remember how well that rebadged Astek unit did, I'd have to find some numbers to refresh my memory.

 

[Rob Williams]

What temps was this occurring at?

Behold my 4.2GHz overclock.



I told you I have cooling issues.

I so know it's in there or available as an option, because some other vBulletin forums resized images for me. I just now made sure it was vBulletin and not phpbb.

Oh, I know it's there. I just can't find it. I'll have another look soon.

According to the guy behind Real Temp some users reported throttling at 80c, some don't report throttling until 85c or even higher.

It's hard for me to tell because the temps go high fast, but I do believe mine throttles at around ~85C.

 

[Kougar]

Behold my 4.2GHz overclock.



I told you I have cooling issues.

LOL wow, that image is worth it just to see RealTemp's Thermal Status: HOT reading. I've never seen that before even on XS.

It's rather odd since a TRUE at 1.41v shouldn't get that hot, but if you are using 1.45-1.5v and not using a 1600 or higher RPM fan then I could see it happening. It really looks like yout got a hot chip, not even a minute into load, its throttling hard, and temps are still climbing... just wow. What fan are you using, anyway?

 

[Rob Williams]

I did another quick test, to see at which point the CPU throttled. I took a screenshot the moment I saw the clock speed decrease, and this is what I got:

For whatever reason, the chip here throttles even before it hits 80C, unless that counter just didn't update as fast as the clock speed.

As for a fan, I'm using the FM123 from SilverStone: http://silverstonetek.com/products/p_contents.php?pno=fm123&area=

I think I goofed last night though. I accidentally had the fan turned all the way down, which was 900 RPM. I just tested again with full speed (2600 RPM), though, and that's what I got that above screenshot with. It doesn't seem to be able to hit 100C with this fan speed, but it is at 96C currently. Time to stop LinX ;-)

 

[unclewebb]

Rob, the best program you can use to detect any throttling of the multiplier is i7 Turbo. You can download it here:

http://www.fileden.com/files/2008/3/3/1794507/Turbo.zip

Here's the latest version of RealTemp

http://www.fileden.com/files/2008/3/3/1794507/RealTempBeta_330.zip

The Thermal Status area has 3 possibilities. OK means that there have been no thermal throttling since you powered on. LOG means that at least one thermal throttling episode has been logged since you powered up and HOT means, thermal throttling is in progress.

I hope to see some i7 Turbo screen shots that show some throttling in action.

I just heard from Asus today. They are working on releasing some new bios versions for the P6T series so users can disable Turbo throttling if they need to. So far they have the P6T Deluxe V1 available for testing. If anyone needs that, let me know.

Edit: There are 3 items that can trigger Turbo throttling. Core temperature is only one of them. Excessive core voltage or core current can also trigger Turbo throttling. Everry CPU is unique so I don't think there are any exact numbers to define this. It's more a combination. Intel has a formula but I won't hold my breath waiting for them to release it.

When your CPU hits 100C then Thermal Throttling kicks in regardless of what motherboard you're on. The multiplier will start to cycle down rapidly to 12.0. This happens so many times a second that traditional software might miss it. That's why the i7 Turbo tools are handy to keep an eye on things. For best results, avoid using any programs that monitor motherboard voltages like HW Monitor and Everest while testing with i7 Turbo.

 

[Rob Williams]

Welcome to the forums unclewebb, and thanks for the explanations, and also the work on both of these tools! I downloaded both RealTemp 3.30 and i7 Turbo and stressed the CPU once again. Here are the results:



The log file from i7 Turbo is here (right-click, save as). I'm not quite sure I understand the log, but if it's of any use to you, then great.

In the BIOS for the Rampage II Extreme, there's an option for "CPU Turbo Power Limit", which seems essentially to be the Turbo throttling. I disabled it, and sure enough the core clock kept stable for the most part up until it hit 100C. Still, far, far too hot to keep stressing for too long, that's for sure.

 

[unclewebb]

Thanks for the welcome and thanks for posting your log file. It tells the full story far more than any single screen shot can.

The important column is the C column in Excel which has the heading C_MULTI or Calculated Multiplier. The column at the end that shows the Load % is also important.

When you are running a single threaded app, your multiplier jumps up to 27 so your MHz goes up to 4209.57 MHz. If your log file was while you were running LinX, there is about 24 seconds between each round where the multi jumps up to 27.

As soon as the load kicks in and goes back to 100%, the multi immediately drops down to 25. As the temperature goes up and Turbo throttling starts, this slowly sags down lower and lower. The Calculated Multiplier data really shows this info clearly. Once you hit 100C then the regular thermal throttling starts to kick in which cycles your multi many times a second down to 12. It does this so quickly that there isn't a huge drop in performance but it does help keep your CPU from going nuclear.

When you're hitting the thermal throttle, your overall average multiplier is only down to 23.75 which isn't that bad considering you are hitting 100C. If you kept at it and more cores hit 100C and started to throttle as well then your calculated multiplier would drop further.

On 975 chips, your bios controls the 27X multi when lightly loaded but I'm sure you already knew that. You could probably raise your 24/7 overclock if you dropped that a notch or two. Save LinX testing for when you get a better cooler. Your temps are contributing to global warming.

Edit: If you go into your Power Options and set Minimum processor state to a low number like 40%, your calculated multiplier should stabilize at idle. Raise this to 100% and disable C1E to keep it as high as possible at idle.

 

[Rob Williams]

I'm glad to hear that the log can be of some use. If you need me to perform any more robust tests, please let me know. Also, the CPU cooler situation is a bit sad, because as it stands, I'm using a top-of-the-line offering (Ultra 120), so I'm not sure what it's going to take aside from water to improve things (and I can't even see high-end water making a huge difference here).

As for the minimum processor power, something like that might be there, but I'm not sure. At this point I'm not too concerned... I don't want this thing to be constantly overclocked, since we use it for our graphics card testing machine (and soon our motherboard testing).

 

[Merlin]

I imagine you can over clock on almost any board, but what is the main things to change?

I'm back home and missed a few things

 

[Rob Williams]

Wait... you are not overclocked?

On i7, you need to worry about the Base Clock primarily. I'd slowly work that up and test for stability in between each hike (I'd go 10MHz higher each time). The QPI voltage is one you'd want to worry about. I'm able to hit 4.2GHz no problem while leaving the CPU voltage to auto, so hopefully it's the same for you.

 

[Kougar]

I got really cold one evening so I decided to have still another crack at 4.5GHz... having a need for the speed due to F@H's -bigadv program only gave added incentive. I never thought I'd find the limit where piling on more vcore didn't help, but I did. Unfortunately 4.4GHz is the absolute max STABLE I can run this individual CPU at, and only when the room is chilled.

When overclocking a CPU there becomes a point where raising the frequency requires an increasingly disproportional amount of voltage to make it stable, and this point occurs right on the edge of what the quality of the silicon can handle. For my chip 4.3GHz runs at about 1.43v (going by CPUZ), 4.4GHz runs at ~1.46v, but if the room isn't chilled it will eventually power cycle despite not even reaching the 100c throttling threshold. By comparison 4.2GHz needed ~1.36v for stability (I'm not sure what the CPUZ reading was as I used to ignore this).

I'm not getting into chilled waterchooling (although I have dabbled with it in the past), and despite the strong initial temptation I'm not going to plunk down the cold hard dinero for one of these beauties. It occurred to be I'd get much better results putting that $380 into a new die-shrunk 32nm Gulftown....

Still, until then I'll be folding proteins with the few extra MHz... each 100MHz boost saves 2 hours off the total calculation times per project.

 

[Rob Williams]

I'm not getting into chilled waterchooling (although I have dabbled with it in the past), and despite the strong initial temptation I'm not going to plunk down the cold hard dinero for one of these beauties. It occurred to be I'd get much better results putting that $380 into a new die-shrunk 32nm Gulftown....

Yes! I'd agree that picking up a Gulftown and overclocking that would be the best solution to your problems... just imagine adding 50% more performance to what you have now ;-)

 

[Kougar]

Yes! I'd agree that picking up a Gulftown and overclocking that would be the best solution to your problems... just imagine adding 50% more performance to what you have now ;-)

Well what is amusing is I probably would only cut the 2.1 day completion time down by a fourth to a third, but it would be rather welcome!

I have to say I can't regret more not buying a second kit of RAM now that I need one... If I did upgrade to Gulftown I absolutely could not avoid having to buy a second kit. Prices on RAM identical to mine have tripled, and it is completely absurd given what it was selling for a year ago! If I'd known I could have ordered a pallet and sold it for a 250%+ return on investment. Damn I just hate hindsight sometimes....