Overclocking Retail Intel Core i7-5960X CPUs

We sit down with Intel's Paul Zagacki to talk Devil's Canyon, Haswell-E and overclocking. Then, we overclock five retail Core i7-5960X CPUs to their limits.

Introduction

Intel’s Broadwell architecture was introduced in early June as the fifth-gen Core processor family (if you missed our coverage, check out Broadwell: Intel Core i7-5775C And i5-5675C Review). Skylake, the sixth-gen design, is expected soon. But if you’re a power user running really well-threaded applications, then both of those “mainstream” platforms remain secondary to Haswell-E—the beefiest incarnation of Intel’s fourth­­-gen architecture

To be sure, Haswell-E remains the enthusiast's top choice almost a year after its introduction. There’s the fact that you find it in six- and eight-core configurations, of course. Huge 15 and 20MB last-level caches are also great for performance. A quad-channel DDR4 memory controller serves up unprecedented bandwidth. And, depending on the model, you get 28 or 40 lanes of PCIe 3.0 connectivity. Oh, and don’t forget that all three models—the Core i7-5820K, -5930K and -5960X—sport unlocked multipliers for overclocking.

As a matter of personal preference, I tend not to spend a bunch of time overclocking when a CPU is first introduced. There’s the fact that the samples typically come directly from Intel or AMD, motherboard BIOSes aren’t particularly mature and, well, it’s just a long loop of trial and error ahead of time-sensitive launch coverage. But during a recent conversation with the folks at Intel, the subject of Haswell-E and overclocking was brought up. I realized that Tom’s Hardware hasn’t done much with Intel’s highest-end platform since last August, and I had an idea for an experiment.

Overclockers' Concerns

Actually, the discussion started with Devil’s Canyon, which takes us even further back to Core i7-4790K Review: Devil's Canyon Tantalizes Enthusiasts. The Core i7-4790K was the processor Intel introduced after causing consternation amongst power users when they saw Haswell's lackluster overclocking. Among the improvements cited in our review, a more efficient thermal interface material and additional capacitors to stabilize power delivery were the most notable.

Underneath that superficial summary, which in no way conveys the effort that goes into bringing a new product to market, I knew that Intel had its own engineers working on characterizing the -4790K’s frequency scaling in such a way as to facilitate more aggressive stock clock rates. It also needed to substantiate the claims it was making about a significant population of processors hitting 5GHz. We didn’t get that far in our review, though we did use a conservative maximum voltage and still had quite a bit of thermal headroom in our sample. I wanted the inside story, straight from the horse’s mouth.

Devil's Canyon Overclocking

And so Intel’s hooked us up with Paul Zagacki, principal engineer with the company’s client computing group, who walked us through some of what the team went through in bringing Devil’s Canyon up and subsequently optimizing the platform for higher stable clock rates.

Core i7-4790K Temperatures

During the course of our talk, Paul did hit us with a tip that we were quick to put to good use: around 80 °C is where the processor starts encountering a roll-off—it’ll actually overclock better under 80 degrees than if you let it continue climbing.

Devil’s Canyon Lessons

The idea came up to revisit Devil’s Canyon using retail samples. It seemed like most of last June’s Core i7-4790K coverage saw reviewers landing around 4.7 or 4.8GHz with four cores under load. Might refinements made over the last year change that? And what about the impact of boxed parts rather than our pre-production CPU?

Haswell-E is going to endure beyond the launch of Skylake, though. It made more sense to do an overclocking test using Core i7-5960X instead. Plus, the -5820K, -5930K and -5960X all use solder as an interface material instead of a polymer. It dawned on me that we never really asked Intel ahead of the Haswell-E launch whether those parts benefited from the lab work gleaned from Devil’s Canyon. I followed up after our initial chat with a series of email questions, which Paul answered in more depth than we often get from Intel:

Tom’s Hardware: Did Intel use any of the experience that went into Devil’s Canyon to improve overclocking on Haswell-E?

Paul Zagacki: Yes. In fact, we used many of the learnings from the Devil’s Canyon analysis and optimization to improve the overclocking experience of HSW-E. With double the core count, -5960X’s total consumption under extreme overclocking can be very high. To ensure those four extra cores scaled well, we allowed additional power delivery by uncapping Icc Max. We also added a knob to allow enthusiasts to increase the thermal throttling point (we didn’t want a few degrees to get in the way of that last bin). This is something we explored for Devil’s Canyon, but found the throttling knobs weren’t limiting, so they were never released into the wild.

Tom’s Hardware: Was the impact any different on the larger HSW-E die compared to quad-core Haswell?

Paul Zagacki: In general, a larger die area benefits single-threaded thermals with more area under the IHS to spread the power. But all cores overclocked becomes more dependent on overall effectiveness of the thermal solution to bring the extra power out of the package. All of those cores running can quickly become power delivery-limited. Power = CV^2F (more or less) and is a cruel mistress, especially when it takes any extra voltage to get the frequency. Adding cores linearly scales the switching capacitance for any given workload.

Tom’s Hardware: What other knobs don’t get enough attention when you see enthusiasts trying to maximize an overclock or dial in stability?

Paul Zagacki: As you know, the extreme overclockers leave no stone unturned. But casual overclockers don’t seem to play with the real-time knobs accessible via Windows. Many of Intel’s overclocking settings can be changed live within the OS. Overclockers can improve their reliability by dynamically changing frequency, voltage, power limits, etc. Intel’s Extreme Tuning Utility, for example, offers a streamlined interface for making such adjustments. It’s a handy tool for overclocking and has some options to dynamically apply different profiles based on the applications running on your system.  For example, XTU can dynamically overclock the processor when a specific game starts up. On top of that, we have put a lot of thought into how to recover a system when you’ve gone off the deep end with a set of overclocking knobs. There are a number of features within the CPU and chipset that sense when we are hung and can automatically reset the platform to a place where a simple recovery to a safe operating point is possible. 

Tom’s Hardware: With all of this talk about Devil’s Canyon and Haswell-E, what are the top overclocks you’ve seen in Intel’s lab?

Paul Zagacki: On Core i7-4790K, we were seeing a strong population of parts reaching 4.9 and 5.0GHz for 4C/8T on water cooling. We have the ability to look at a lot more silicon than your average overclocker, and then can correlate those results back to data we see in manufacturing and test to project against huge populations of data. Our focus for Devil’s Canyon was on what a typical enthusiast would do in this space, generating stable results and focusing less on things like getting the best-case frequency snapshot, single-threaded or LN2.

Core i7-5960X overclocking has also been amazing. Even with eight cores, the top clock rate is only slightly lower than the four-core -4790K, with some units clocking in as high as 4.7-4.8GHz. This is pretty amazing considering all the additional power/temperature and resulting voltage drops you have to manage with 4x the switching capacitance from the additional cores. Also, we’ve also seen DDR4 overclocking above 4,000 MT/s. This is truly amazing—DDR4 was just released and the officially supported maximum frequency is 2133. With all those cores running, you have to feed them, so the DDR4 overclocking helps you realize a lot of performance you gain from tuning your cores.

How We Tested

Not long after our talk, a box showed up with five Core i7-5960X CPUs in sealed retail packaging. I reached out to the folks at NZXT for a reasonably-priced closed-loop liquid cooler and, a few days later, was greeted by a Kraken X41 at the office.

Test System Components










MORE: Best Gaming CPUs For The Money
MORE: How To Build A PC
MORE: All CPU Content

Test Results

Over the course of a weekend, we got all five CPUs tested, recording their top stable eight-core clock rate, single-core clock rate, maximum full-load temperature, peak single-core temp and the voltage required to get each sample stable for an hour of RealBench. Along the way, we took a number of interesting notes.

Let’s start with the big number—top clock rate with all eight cores fully utilized. In four out of five cases, we hit a ceiling at 4.5GHz. The fifth made it to 4.6GHz. That’s not to say those are absolute maximums, though. As Paul mentioned, a full load taxes the thermal solution’s ability to move heat away from the IHS. NZXT’s Kraken X41 is a great closed-loop cooler for around $110, but it’s not the largest or most aggressive of its kind. And even then, given the complexity of Haswell-E (a 2.6 billion-transistor die) compared to Haswell (at 1.6 billion, much of which is integrated graphics), 4.5GHz is a strong increase.

Working within the bounds of our cooler, we started tuning the other direction—find the top clock rate, flirting with voltages up to around 1.4V, and then back the voltage down to the edge of stability for a lower temperature.

Knowing that 80 degrees was where overclocking began to roll off, it became immediately apparent that 1.4V was too much for our cooling setup. So, most serious attempts started around 1.35V. From there, we’d back off in coarse .1V increments before adding voltage back in .025V steps at the first sign of instability. And although there’s no clear correlation between our top overclock and the voltage it took to get there, there is some consistency in the temperatures we saw.

Generally, the CPUs that were stable at 1.2V encountered the lowest temperatures. We’re reporting readings from the hottest core—meaning seven others ran cooler. Sample 5 looks like a sweet little piece of silicon. And although Sample 4 is the one that ran stably at 4.6GHz using a 1.3V Vcore setting, its hottest core is, well, five degrees hotter than what we see from any other processor.

Single-core overclocking was a little trickier, even though the Kraken had a lot less thermal energy to contend with. As you can see in the image above, Windows likes to schedule single-threaded workloads on different cores. So, your first five loops at 5GHz might be successful, only to crash on the sixth as a weaker core buckles under the load. Not surprisingly, then, it’s a real pain to loop something like our iTunes or LAME workload over and over looking for instability.

But after many hours of testing, we’re pretty confident that these chips could all take 4.8 and 4.9GHz across their eight cores, so long as only one was active at a time.

Depending on the core under load, temperatures were all over the place. We kept our eye out for the warmest-running one and made it representative of that CPU.

A Sign Of Things To Come

I’ve made no secret of my disappointment with Ivy Bridge and Haswell. After the gloriousness that was Nehalem and then Sandy Bridge, the following two architectures made me question Intel’s commitment to the desktop PC.

The expediency with which Intel made Devil’s Canyon happen after receiving the community’s feedback on Haswell signified a change in direction, though. And while we haven’t gotten our hands on a sample capable of 5GHz yet, Intel says it’s seeing a notable population of parts capable of that symbolically significant clock rate.

Instead, we were able to test a quintet of something even juicier—the eight-core i7-5960X. Retail examples all, we had little trouble taking all five to 4.5GHz (a 1.2GHz overclock over the chip’s 3.3GHz stock Turbo Boost setting with eight cores active). Single-threaded frequencies bounced between 4.8 and 4.9GHz. Despite Haswell-E’s business-class pedigree (these are the same dies used to make Xeon CPUs), Intel exposes a handful of additional tuning options to make the Core i7s even more attractive to enthusiasts. Needless to say, we're happy to see it.

MORE: Best Gaming CPUs For The Money
MORE: How To Build A PC
MORE: All CPU Content

Chris Angelini is a Technical Editor at Tom's Hardware. Follow him on Twitter and Google+.

Follow Tom's Hardware on Twitter, Facebook and Google+.

This thread is closed for comments
53 comments
    Your comment
  • Bossyfins
    5 5960xs... What the hell O.o
  • blackmagnum
    If you have money to spend on an i7-5960X, you probably have enough to pay someone to tune it (overclock) for you.
  • spentshells
    I would have preferred a very high end cooler...despite the fact we seemed to already hit the performance limit of these chips.
  • iam2thecrowe
    Quote:
    If you have money to spend on an i7-5960X, you probably have enough to pay someone to tune it (overclock) for you.

    where's the fun in that?
  • wtfxxxgp
    I think the results are impressive. Wonder what you would have gotten had you gone gung-ho on the cooling. An 8-core CPU that can overclock like this is a beast for any power user.
  • tsnor
    I REALLY liked hearing intel give the OC distribution for the i7-4790K, if there is any possible way you could give the numbers for the i5-4690k that would be great. Nice article.

    If you still have the parts an update with the most extreme cooler you could find against one of the CPUs would be interesting, so would freezing yourself (dropping the ambient) and seeing if that allowed another 2 bins, or if it was rounding error.

    I missed seeing system power @ the OC frequency, it might be in there tho, re-reading to see if i missed it.
  • mercsniper
    any post on the other numbers used in these overclocking tests? I've had problems keeping my system stable at 3.8Ghz and DDR4 at 2400MHz. EPU was turning on and not turning off causing significant strain and requiring a reboot. I ended up turning EPU off.
  • Math Geek
    i'd also like to know the power draw from such an overclock. i'm sure it was a lot more than the 140w tdp at stock.
  • Daniel Ladishew
    I would also like this set of tests conducted using a premium cooler to show if the significant increase in 5.0Ghz chips is represented in any of the 5 you tested. I can hit 4.5Ghz on my i5 4690K, so seeing 4.6Ghz from the "better" chip does not really show anything worthwhile to me. I think we would all like to see exactly why this chip is worth 1K+ from an overclocking standpoint.
  • Bossyfins
    Well, it could show Intel wanted to make sure everyone could atleast have a decent cooler. Also, recently on WCCFtech, sources shoe skylake hitting 5 on air.
  • vertexx
    Nice article. I'm very interested in a similar one for the 5820k, especially interested whether the 5820k would produce less thermal output under overclocked load.
  • TechyInAZ
    Wow, 8 cores at 4.8ghz! That's insane. Very cool stuff.

    Too bad it's still $1000, I think $750-$800 should be it's best price point.

    Quote:
    Quote:
    If you have money to spend on an i7-5960X, you probably have enough to pay someone to tune it (overclock) for you.
    where's the fun in that?


    I can see your point, why would you want to risk damaging $1000 of CPU for a overclock? Well I defiantly do not recommend overclocking a CPU like this if your a newbie (if you really want to though, read a TON on overclocking haswell-E), but the experienced guys who know exactly what they're doing can easily overclock any CPU without destroying it. ;)
  • 10tacle
    Nice to see there is still a decent amount of headroom at 1.2GHz with this new chip, which is a whopping 50% better than the 800MHz increase I have my 4690K at over stock boost: 4.7GHz. And that's a good overclock for that chip while remaining stable and cool.

    This is just 100MHz shy of the overclock leap my Sandy Bridge 2500K was able to knock out with a slower stock turbo boost speed of 3.5GHz and overclocking to 4.8GHz (I did crack 5GHz successfully with it on four cores, but for longevity backed it down).

    As one who has been enjoying overclocking since the Celeron 300A->450MHz days for an easy 50% overclock, I understand with continued µ shrinkage we are approaching the limits of physics and thermals. A good successful overclock today is between 25-35% over stock turbo boost. I see those figures getting smaller. The days of an easy 50% overclock with off-the-shelf cooling are long gone.
  • red77star
    LGA 2011-V3 is a platform to get. i7 5820k is nice entry CPU. LGA 1150 is waste of money in my opinion.
  • TechyInAZ
    202972 said:
    Nice to see there is still a decent amount of headroom at 1.2GHz with this new chip, which is a whopping 50% better than the 800MHz increase I have my 4690K at over stock boost: 4.7GHz. This is just 100Mhz shy of the overclock leap my Sandy Bridge 2500K was able to knock out with a slower stock turbo boost speed of 3.5GHz and overclocking to 4.8GHz (I did crack 5GHz successfully with it on four cores, but for longevity backed it down). As one who has been enjoying overclocking since the Celeron 300A->450MHz days for an easy 50% overclock, I understand with continued µ shrinkage we are approaching the limits of physics and thermals. A good successful overclock today is between 25-35% over stock turbo boost. I see those figures getting smaller. The days of an easy 50% overclock with off-the-shelf cooling are long gone.


    Yep. The glory days of Sandy Bridge are over (Sandy bridge is still my all time favorite architecture from Intel).
  • joex444
    For something like this if you're looking for a CLC, why would you settle with a single 140mm radiator? The Kraken X61 is a 280mm part which should let it handle a meaningful amount of extra heat. Intel's giving 4.7GHz as an example and here we see 4.5GHz with what I can only think is a small radiator for such a task.

    I'd really be interested in overclocks for the other two LGA 2011-3 Core i7 parts. The 5820K is by far the most affordable and has fewer PCIe lanes, while the 5930K has a higher base clock and the full 40 PCIe lanes. Given both of these parts are 6 core one would assume they could overclock better.

    Also fascinating would be actual benchmarks. What does 4.5GHz on all 8 cores actually look like? How does that compare to a stock 5960X, 5930K, or 4790K?
  • gamebrigada
    Got a 5960x last year at PAX, was the winner of the intel badge thingie. I can almost say it wasn't worth it, the financial burden to make a machine around it was spectacular at the least.

    Anyway, running at 4.5 stable with a kraken x41. I dropped it down to 4.4 because it would at times crash, but I believe it to be a motherboard issue (x99 deluxe from asus) rather than the proc. These procs are unbelievable, but I'm thinking intel has gone a long way at getting their shit together, when I started with the one I got, the maximum overclocks on these puppies was all over the board. anywhere from 4.3 to 5.0.
  • ralanahm
    My peaks 250 to 300 watt according to Asus 1 second average. Using big nocturnal with three 150mm fans on full. But I am not using the advanced/ln2 OC settings just normal and for long runs it trottles 80c and keeps the watt in 220 to 250 range.
  • alidan
    Quote:
    Quote:
    If you have money to spend on an i7-5960X, you probably have enough to pay someone to tune it (overclock) for you.
    where's the fun in that?


    you need the power than 5ghs can enable but don't have the ability or knowledge to do it yourself, so instead of risking the processor not knowing what you are doing, pay someone who does to push it to its limits.

    granted, single core performance will get the biggest benefit from this, all the heavy loads i would want to put on an 8 core cpu would be better served by a multi cpu motherboard and getting a few 6 core chips for less than 1 8 core cpu
  • adamdbz
    I'm still rocking an I5 2400.
    only game that made me want to get a better CPU was Gta V
    with an r9 290 on 4480x 1024.
  • knowom
    Quote:
    Quote:
    Quote:
    If you have money to spend on an i7-5960X, you probably have enough to pay someone to tune it (overclock) for you.
    where's the fun in that?
    you need the power than 5ghs can enable but don't have the ability or knowledge to do it yourself, so instead of risking the processor not knowing what you are doing, pay someone who does to push it to its limits. granted, single core performance will get the biggest benefit from this, all the heavy loads i would want to put on an 8 core cpu would be better served by a multi cpu motherboard and getting a few 6 core chips for less than 1 8 core cpu
    As would single core performance if they actually made CPU's and boards for that purpose.

    That's what AMD should focus on single and dual core APU's with great IPC on a dual socket platform that can dual/triple/quad crossfire the APU graphics with each other along with discrete graphics.

    Not only that, but dual socket enables the possibility of more memory capacity as a nice bonus and better cooling dissipation due to two heat sinks which is a nice bonus.
  • Math Geek
    now that is a very intriging sounding system. piggybacking the igp's from multiple APU's could make for some very interesting potential capabilities. modular mobo that hold 4 of them allowing a single one to start and adding as needed. hhhhhhhmmmmmmmmmmmmmmmm

    *checking phone book for amd's #*
  • Calculatron
    Glorious!
  • Bartendalot
    Quote:
    LGA 2011-V3 is a platform to get. i7 5820k is nice entry CPU. LGA 1150 is waste of money in my opinion.


    For someone just building a gaming rig?