Sign in with
Sign up | Sign in
The 5 GHz, Six-Core Project: Core i7-980X Gets Chilly
By ,
1. Cooling Comes Full Circle

Our history with sub-ambient cooling goes back to the early days of this site, with our discussion of Kryotech’s first phase-change cooler in 1997 and our test of the improved version in 1999. Our 618 MHz Celeron quickly gave way to 800 MHz and 900 MHz Athlons, and even more cooling revisions pushed us past the GHz barrier less than a month later.

And then the competition showed up. Before Asetek became the purveyor of low-cost liquid cooling, its VapoChill phase-change system streamlined case and cooling components to a single box that supported processors from both AMD and Intel. By the time Prometeia threw its hat into the ring Kryotech had vanished and we had blown past 3 GHz. A few more improvements got us to 4.1 GHz, but reaching the next level would require a major improvement in either CPU technology or cooling capacity.

Liquid nitrogen cooling on the CPU and a phase-change cooler on the chipset finally allowed us to go beyond 5 GHz six years ago, an effort that inspired competitive overclockers the world over. Of course it wasn’t practical, because there were no practical 5 GHz solutions at that time. It wasn’t until Intel released CPUs on its 32 nm production process that “permanent” cooling solutions began to look viable for achieving these speeds. In fact, Chris Angelini was able to boot at 4.93 GHz using his Core i5-655K on air.

The technology was now in place to forgo temperamental liquid nitrogen cooling in our 5 GHz efforts, yet reaching this speed on a low-cost CPU that was already capable of running stably at 4.6 GHz with air cooling would have appeared trivial. The performance of a 5 GHz dual-core in a world of quad-cores would have been likewise laughable. We needed a properly high-profile, high-tech CPU to make this article worthwhile.

We knew we had a the perfect start for our project when Intel finally released its six-core, twelve-thread Core i7-980X Extreme Edition in March. A cooler with adequate capacity would still be required, and FrozenCPU.com matched us up with the hardware

2. The Compressor Returns

Whether the knowledge comes from a school science class, television, or the Internet, many of us already know the basic principles of phase-change cooling. Gasses absorb heat during expansion, cooling the surrounding area. A gas that has been compressed to a liquid state absorbs even more heat as it changes phase back to a gaseous state, in the same way that water absorbs a great amount of energy as it boils. Gasses likewise give off heat during compression, which is why traditional refrigerators and air conditioners use an outside radiator to remove heat from the gas after it’s compressed. Cooling the compressed gas allows it to change to a liquid state, giving way to the term “condenser” as a name for the “hot-side” radiator.

The evaporator is where a CPU phase-change cooler departs from a traditional refrigerator or air conditioner. While the gas-to-gas heat exchangers of those familiar household appliances resemble a second radiator, a CPU chiller uses a much smaller block-shaped evaporator to draw heat away from the CPU.

What appears to be a chunk of copper at the end of a hose on the unit above is actually a hollow evaporation chamber connected to two high-pressure lines, with the higher-pressure side delivering the liquid and the lower-pressure side drawing away the resulting gas. The manufacturer for this device, Cooler Express is a Taiwanese firm that produces a variety of single-head, dual-head, chilled liquid and cloud-chamber coolers for electronics cooling, production processes, and laboratory research environments.

The list of parts included in the package will mostly be determined by the seller. Because the original bracket set did not support Intel’s most recent processors, FrozenCPU.com had a new mounting block machined out of aluminum to support LGA 1156 and LGA 1366 sockets. The aluminum mount is anodized black to match the original plastic part, and also includes a new socket support plate with corresponding holes. FrozenCPU.com adds the remaining hardware to complete the second mounting kit, allowing builders to borrow parts from the original mounting kit if needed. FrozenCPU.com sells the complete unit with both installation kits as its Cooler Express 2010 Super Single Evaporator CPU Cooling Unit.

3. The Test Platform

Intel’s Core i7-980X six-core processor requires an LGA 1366 motherboard with recent BIOS, and we just reviewed a few of those in May. Of the boards we evaluated, Gigabyte’s X58A-UD7 appeared to offer the best stability.

Testing the performance differences of a single processor at various speeds wouldn’t require fast RAM, but we used it anyway.

High-end graphics also would not be required, yet we didn’t want the system to appear slow in game benchmarks. Sapphire’s Radeon HD 5850 is more than adequate for producing playable frame rates in most games at high settings.

The Cooler Express startup guide suggested a motherboard installation height that almost perfectly matched our Danger Den Torture Rack 2 chassis. That’s a striking coincidence, since we were already using this chassis in a motherboard testing station.

Corsair’s reputable CMPSU-850HX was already installed on the Torture Rack 2 from its use in the motherboard testing station, and we knew this unit would provide far more power than our single CPU and graphics card would require.

4. Cooler Express Installation, By-The-Book

The Cooler Express CPU chiller comes with an installation guide that’s intended for sockets that either don’t have a rear support plate, or in the case of AMD, have had the factory support plate removed. FrozenCPU.com-supplied LGA 1366-compatible hardware substitutes in for the original pieces.

Four mounting screws thread through the back plate, with four plastic spacers added to prevent crushing of the included insulation sheets. These spacers probably aren’t long enough to fully serve their purpose when installed over the socket support plates of LGA 1156 and LGA 1366 interfaces, but we made no changes here.

A thick piece of insulating foam is then placed over the support plate, following by a heating element. The heating element prevents condensation from forming on the back side of the motherboard when the CPU chiller cools the socket.

The instructions said nothing about adding another sheet of insulation, but the unit was designed to be mounted over a flat surface. We cut an additional piece of thin foam to fit around the socket’s original support plate, closing the gap between the motherboard and the original insulation sheet.

The cooler’s support plate screws are then inserted through the motherboard’s original CPU cooler mounting holes. Four nuts secure this support assembly, using plastic washers to prevent motherboard scratches and related circuit damage.

5. Insulation Installation

Continuing our by-the-book installation, we added two sheets of insulation around the CPU socket, trimmed to clear onboard devices. Notice that there is no space at all for insulation between the CPU socket and one side of the CPU voltage regulator.

Having no insulation to seal one side is like “asking for trouble” from condensation, but we nonetheless continued to follow the installation manual by taping over air gaps around the socket’s top.

A final insulation barrier is meant to seal the top of the CPU socket to the bottom of the evaporator’s mounting block, but it looks a little thin to us.

One of the problems we see is that the final foam sheet (above) is thinner than the portion of the evaporator protruding through the mounting block (below).

6. Just Add...Water?

Somewhere between the previous page and this page we brushed a layer of Zalman’s ZM-STG1 thermal grease onto the top of our CPU.

Four springs fit over the support assembly’s screws, sliding into recesses on the mounting block. These are topped by four more plastic washers.

Four nuts compress the mounting springs, allowing the unit to self-center the evaporator’s pressure against the CPU.

The Cooler Express manual reminds us that it’s important to dry the motherboard at the end of every day. Hoping to get through a day’s testing before enough water accumulated to short-circuit our CPU or motherboard, we fired the system up and began testing. Around four hours passed before the system started throwing wild errors.

Several of the processor’s contacts were darkened, but the system still worked. We obviously needed to take greater precautions than those recommended by the installation guide.

7. Reworking The Installation

We’ve already proven that we’re not afraid to risk a little “free hardware” to prove a point, and we’re completely familiar with methods that would have prevented condensation from accumulating in fragile areas. These methods include painting the entire board with nonconductive sealant, sealing the entire area around the CPU with nonconductive putty such as kneaded eraser, and filling the LGA with dielectric grease. Unfortunately, most of these solutions cannot be completely removed from the motherboard. While we only needed our system to run for 12-hour intervals, we do recommend most of these precautions for extended use.

The one drop of water that stopped the system had left a trail, starting at the evaporator and running past the round hole in the foam barrier, down the side of our CPU socket, and into the Land Grid Array (LGA). Nonconductive putty was added to fill the gap between the CPU’s heat spreader and socket’s pressure plate.

A new layer of tape seals the CPU area to the reinstalled, custom-cut foam layers. The foam sheet with the round hole was reinstalled over this tape.

A bead of putty fills the space between the evaporator’s mounting block and top foam sheet. These minor changes allowed our system to run eight, but not the full twelve hours between dry-offs.

Removing the CPU after an eight-hour test session revealed drops of condensation on each of its 1366 contact pads. Air circulating under the CPU was the problem, and that problem can be solved by filling the LGA with dielectric grease. Petroleum jelly makes an adequate substitute for dielectric grease when used at moderate to low temperatures, such as those experienced with sub-ambient CPU cooling systems.

As seen in our 2008 Overdrive Competition, applying putty around the CPU socket is another option to further prevent air from getting underneath it. Unfortunately, voltage regulator chokes immediately adjacent to the socket of the motherboard we used today would have severely obstructed such work.

8. Basic Overclocking

The Core i7-980X might be labeled as a 3.33 GHz model, but Intel's Turbo Boost technology automatically increases its clock rate to either 3.60 GHz or 3.47 GHz, depending on the number of cores being used. The screen shot below doesn’t show exactly 3,466 MHz because Gigabyte doesn’t set its base clock to a proper 133.33 MHz. With a default base clock of around 135 MHz, choosing manual settings to reach Intel’s intended 133.333 MHz base clock results in an actual clock speed of 133.0 MHz.

Notice that the highest temperature reached by our CPU at stock settings was measurable, exceeding the lowest-possible reading of its DTS (digital thermal sensor) at -12° Celsius. While we couldn’t track temperatures this cold in real time, the fact that they occasionally exceeded the minimum threshold tells us that the CPU core is much warmer than the -50° evaporator temperature reported by the Cooler Express status monitor.

I dropped my personal test voltage limit for Intel's 32 nm technology to 1.35V after a rash of blown processors swept Tom’s Hardware labs at settings as low as 1.375V. But those articles used air cooling, and processors do gain some voltage tolerance as temperatures are lowered. Today we use my air-cooled 1.35V limit as the starting point for our chilled-overclocking effort.

A modest 1.35V is already pushing our CPU temperature into positive numbers at full load, and we’re beginning to wonder whether surface imperfections on the evaporator will need to be addressed if we’re to maintain useful temperatures going forward. The CPU is coping admirably, reaching 4.46 GHz at 100% stability.

Our next step is 1.45 volts, which yields a stable 4.69 GHz at full load with Intel Hyper-Threading turning our six physical cores into twelve logical cores. Temperatures are still acceptable, climbing only a few degrees from our 1.35V results. Normal caution would have us stop here, but we’re committed to finding this processor’s limit or reaching 5 GHz, whichever comes first.

9. Reaching The Goal

A little more testing revealed 4.9 GHz at 1.60V and new thermal stability limits. Even the slightest increase voltage would cause our CPU to reset at 39° Celsius, and even the slightest increase in speed would lower the temperature at which our CPU reset. We had to get our temperatures down.

Reports from several cooling sites show that the latest compounds can reduce temperatures by up to 4° compared to Zalman’s ZM-STG1, yet our testing showed that a drop of at least 10° would be required to reach full CPU stability at 5 GHz. The logical next step would have been to smooth the evaporator’s mating surface through processes such as precision sanding and lapping, but we weren’t certain how thick this surface was and didn’t want to risk damaging a borrowed product.

We finally reached our 5 GHz goal not by increasing thermal transfer speed, but by reducing heat. Disabling Intel's Hyper-Threading technology dropped us from twelve logical to six physical cores, while simultaneously dropping CPU temperatures by nearly 20°.

The Core-i7 980X has completely unlocked multipliers, yet no multiplier exists to reach 5 GHz at the stock 133 MHz base clock. We also wanted to retain a consistent memory data rate to assure accurate assessment of CPU performance. Increasing our base clock from exactly 133.333 to 166.666 MHz would have been a perfect solution, using a CPU multiplier of 30 and a DRAM multiplier of 4 to reach 5.00 GHz and DDR3-1333. Gigabyte doesn’t use fractional base clock frequencies however, so the actual base clock was increased from 133.0 to 166.0 MHz.

10. Test Settings
Test System Configuration
CPUIntel Core i7-980X Extreme LGA 1366, 3.33 GHz, 12MB L3 Cache
MotherboardGigabyte GA-X58A-UD7, Intel X58 Express Chipset, LGA 1366
RAMKingston KHX16000D3ULT1K3/6GX (6GB), DDR3-2000 at DDR3-1333 CAS 7-7-7-20
GraphicsSapphire Radeon HD 5850 1GB
725MHz GPU, GDDR5-4000
Hard DriveWestern Digital RE3 WD1002FBYS, 1TB
7,200 RPM, SATA 3Gb/s, 32MB cache
SoundIntegrated HD Audio
NetworkIntegrated Gigabit Networking
PowerCorsair CMPSU-850HX 850W Modular, ATX12V v2.2, EPS12V, 80 PLUS Gold
Software
OSMicrosoft Windows 7 Ultimate x64
GraphicsAMD Catalyst 10.5
ChipsetIntel INF 9.1.1.1020
Benchmark Configuration
3D Games
CrysisPatch 1.2.1, DirectX 10, 64-bit executable, benchmark tool
Test Set 1: High Quality, No AA
Test Set 2: Very High Quality, 4x AA
DiRT 2 DemoIn-game benchmark
Test Set 1: High Quality Preset, No AA
Test Set 2: Ultra Quality Preset, 4x AA
Call of Duty: Modern Warfare 2Campaign, Act III, Second Sun (45 seconds FRAPS)
Test Set 1: Highest Settings, No AA
Test Set 2: Highest Settings, 4x AA
S.T.A.L.K.E.R.: Call Of PripyatCall Of Pripyat Benchmark version
Test Set 1: High Preset, DX11 EFDL, No AA
Test Set 2: Ultra Preset, DX11 EFDL, 4x MSAA
Audio/Video Encoding
iTunesVersion:9.0.2.25 x64
Audio CD (Terminator II SE), 53 min
Default format AAC
Handbrake 0.9.4Version 0.9.4, convert first .vob file from The Last Samurai (1.0 GB) to .mp4, High Profile
TMPEGEnc 4.0 XPressVersion: 4.7.3.292
Import File: Terminator 2 SE DVD (5 Minutes)
Resolution: 720x576 (PAL) 16:9
DivX Codec 6.9.1Encoding mode: Insane Quality
Enhanced multithreading enabled using SSE4
Quarter-pixel search
XviD 1.2.2Display encoding status = off
MainConcept Reference 1.6.1MPEG2 to MPEG2 (H.264), MainConcept H.264/AVC Codec, 28 sec HDTV 1920x1080 (MPEG2), Audio: MPEG2 (44.1 KHz, 2 Channel, 16-Bit, 224 Kb/s), Mode: PAL (25 FPS)
Productivity
Adobe Photoshop CS4Version: 11.0 x64, Filter 15.7MB TIF Image
Radial Blur, Shape Blur, Median, Polar Coordinates
Autodesk 3ds Max 2010Version: 11.0 x64, Rendering Dragon Image at 1920x1080 (HDTV)
Grisoft AVG Anti-Virus 9.0Version: 9.0.663, Virus base: 270.14.1/2407, Benchmark: Scan 334MB Folder of ZIP/RAR compressed files
WinRAR 3.90Version x64 3.90, Dictionary = 4,096 KB, Benchmark: THG-Workload (334MB)
7-ZipVersion 4.65: Format=Zip, Compression=Ultra, Method=Deflate, Dictionary Size=32KB, Word Size=128, Threads=8
Benchmark: THG-Workload (334MB)
Synthetic Benchmarks and Settings
3DMark VantageVersion: 1.0.1, GPU and CPU scores
PCMark VantageVersion: 1.0.1.0 x64, System, Productivity, Hard Disk Drive benchmarks
SiSoftware Sandra 2010Version 2010.1.16.11, CPU Test = CPU Arithmetic / MultiMedia, Memory Test = Bandwidth Benchmark
11. Benchmark Results: 3D Games

We put hours into running our game benchmarks, but to no avail. Our current benchmark set is so GPU-bound that most of the differences aren’t worth showing, let alone discussing. Previous graphics tests have shown that high-end multi-GPU configurations are the only way to break past this barrier at high-quality game settings, but that level of graphics power is not currently available at this lab. Click on any of the images to reveal details at full-scale.

The seventh image above shows the largest game performance difference of less-than 4% occurring at our lowest test settings in the S.T.A.L.K.E.R.: Call of Pripyat benchmark. Beyond that, this CPU is too powerful at stock speeds to show noticeable performance differences.

12. Benchmark Results: Audio And Video Encoding

Moving past the painfully monotonous gaming benchmarks, applications are where the Core i7 shines. Apple iTunes isn’t the best example, with its inability to use every CPU core, but at least it responds favorably to our overclocking efforts.

HandBrake and TMPGEnc show no loss for disabling Intel Hyper-Threading (HT) at 5 GHz, instead handing that configuration a huge 50% performance lead.

MainConcept puts the fastest HT-enabled speed in the lead by 31%, the 5 GHz overclock giving up 12% of its expected performance through the loss of its virtual cores.

13. Benchmark Results: Productivity

Even the CS4 version of Photoshop appears to be limited to six or fewer threads, since disabling HT has no noticeable impact on the 5 GHz configuration’s performance leadership.

One of the first programs we saw to take advantage of four CPU threads, 3ds Max doesn’t appear to use more than six. Our 5 GHz configuration holds its lead at 47% over stock.

AVG benchmark doesn’t work right in Windows 7, and we still haven’t figured out why. At least the 1% difference pointed in the expected direction.

The 5 GHz machine’s performance gain drops to “only” 32% in file compression.

14. Benchmark Results: Synthetics

It’s a little surprising to see minimal performance differences in 3DMark, since this benchmark has a CPU test. The differences seen in the chart below are almost as small as those of actual games!

PCMark appears to have a preference for Intel’s Hyper-Threading (HT) technology, handicapping our highest overclock slightly.

With HT disabled, the 5 GHz overclock appears to be a completely different processor in Sandra, and not in a good way.

15. Power And Efficiency

The Core i7-980X is a fairly effective power miser at stock settings, as our entire system consumed only 104W idle and 202W when running 12 Prime95 threads.

Power consumption begins to look ugly when we go beyond 1.35V core, more than doubling by the time we reach 5 GHz. That can’t be a good sign for an overclock that had a peak performance gain of around 50%.

Average performance gains were far smaller than peak gains, with our fastest configuration pushing a mere 20% over stock.

Intel’s no fool when it comes to picking the clock speeds and voltage levels of its high-end products. Articles as recent as our June System Builder Marathon have shown that efficiency can be improved when an overclocked processor’s performance increases more than its power consumption. But the above charts indicates that our smallest increase in voltage might have been a little bit overzealous for a daily-use machine.

16. Victory At Last?

While our first 5 GHz project helped transform competitive overclocking into a spectator sport, today’s 5 GHz machine achieves some significance for reaching the same speed without the drama that made its predecessors famous. That is to say, today’s project brought a modicum of practicality to the 5 GHz efforts we started six and a half years ago.

Yet, this victory still feels somewhat hollow, and it’s not because dual-core CPUs have already been able to do this for several months. Watching a 5 GHz processor get knocked out by a stock-speed quad-core would have felt like a step backwards. Instead, it’s the remaining impracticalities of pushing a six-core chip to this speed that knocked the wind out of our sails.

First we encountered core temperatures that were too high, in spite of a cooler that never exceeded -40°, and we solved that problem by disabling Intel Hyper-Threading technology rather than addressing the cause. Already feeling that our solution to the heat issue was somewhat lacking, we then noticed that power consumption had increased twice as much as clock speed. Finally, with half the logical cores (on the same number of physical cores) and twice the power consumption, our 50% overclock got us only a 20% average performance gain.

Considering that even our initial 1.35V overclock required a 50% power increase to reach a 25% higher clock speed, it appears that pushing the Core i7-980X by even a moderate amount is wasteful. Those non-overclockers who just breathed a collective “duh” should check out our recent System Builder Marathon, where all three machines increased performance-per-watt by overclocking. Yet, we didn’t try overclocking the Core i7-980X at stock voltage, and we’re certain some builders will find a modest overclock that comes at no cost in efficiency.

Finally there’s the expense. Our $900 cooler requires around 480W of power in addition to that consumed by the rest of the PC. At 12 hours a day of use, that would be 2102 KW/h annually just to run the cooler, worth around $231 at the average U.S. basic residential energy rate. Summer surcharges can double or even triple the cost for anyone whose power consumption exceeds their state’s basic rate limit, a group that includes nearly every electronics enthusiast. Leaving your computer on 24 hours a day to run tasks like Folding@home will double your cost again. If you overclock in an air-conditioned room, the added energy extracted by your air conditioner could double your cost yet again. Thus, while proper system preparation can make phase-change cooling practical in terms of longevity and service life, high cooling cost and moderate performance gains make it a difficult choice to stomach for continuous-duty operation.

Our 5 GHz six-core still makes a great demonstration PC, but the same can be said of any of the liquid nitrogen-cooled systems that have already pushed this same processor model to 6 GHz.