We've already seen what the Core i7-4770K can do at its default frequency. Now, it's time to go beyond the chip's stock speed.
K-series Haswell processors provide overclockers with three ways to crank the CPU frequency. Raising the Turbo multiplier is the easiest path to higher clock speeds. This variable targets the CPU cores exclusively, so changes won't affect other system components. Multiplier tweaking has been the preferred overclocking method for the last few generations of Intel CPUs, and it remains so for Haswell.
The CPU frequency is the product of the multiplier and the base clock speed. Increasing the latter can also produce higher core speeds. However, the processor's DMI and PCI Express interfaces also derive their frequencies from the base clock. Fiddling with the base frequency causes those interfaces to run out of spec, potentially compromising stability.
Haswell supports base clock control, but we're told by multiple folks in the motherboard industry that the range of useful frequencies is similar to Ivy Bridge: an additional 5-10% at best. That said, Haswell improves upon Ivy by adding a base clock strap inherited from Intel's ultra-high-end Sandy Bridge-E processor. This strap acts as a reduction gear for the DMI and PCI interfaces. It's capable of dividing the base frequency by 1.25, 1.66, or 2.5, allowing that clock to be raised to 125, 166, and 250MHz without messing with the chipset and peripheral links.
Well, that's the theory, anyway. Asus tells us it hasn't found a single Haswell CPU capable of running a 250MHz base clock. The majority of chips will do 166MHz, it says, and 125MHz should be a lock for all of them. You should see a similar +/- 5-10% adjustment range at each strap setting.
With Z87 boards supporting CPU speeds up to 8GHz in effective 100MHz increments via multiplier boosting, there's little need to touch the base clock or its associated strap. Only extreme overclockers looking to set benchmark records should worry about those settings.
Asus has tested hundreds of Haswell CPUs as part of its effort to profile the chip for auto-tuning algorithms. According to the motherboard maker, Intel's new hotness has a little less overclocking headroom than Ivy Bridge does. Perhaps more importantly, Haswell apparently has more variance from chip to chip, especially in the voltages necessary to hit specific speeds.
Of the processors Asus has tested, 70% hit 4.5GHz, 30% reached 4.6GHz, and 20% made 4.7GHz. Only 10% were stable at 4.8GHz. Heat is reportedly the limiting factor, and Asus recommends using a dual-fan water cooler to prevent thermal throttling past about 4.5GHz or 1.25V. Going beyond 1.35V is apparently problematic even for high-end water coolers.
Since we have a high-end water cooler in-house, we decided to see how far it could take our Core i7-4770K. This is a different chip than the one Scott used for his CPU benchmarks, and the test configuration differed slightly from his, as well. The system was based on Asus' Z87-PRO motherboard and a GeForce 680 GTX DirectCU II graphics card. Corsair provided the Force GT 120GB SSD, the AX850 power supply, and 16GB of Vengeance Pro memory.
Although the RAM is rated for operation at speeds up to 2400MHz, we confined ourselves to testing the limits of the CPU. Both Asus and Gigabyte tell us that higher memory speeds can limit CPU overclocking, so it may be worth exploring that dynamic in a separate article.
To keep the CPU cool, we used Corsair's H80 water cooler. The radiator isn't a double-wide affair, but it is sandwiched between dual fans. We also swapped the stock spinners for Corsair's Air Series SP120 units.
We kept things simple for this round of tests and limited ourselves to manually tweaking settings via the motherboard firmware. A combination of AIDA64's CPU stress test and the Unigine Nature benchmark was used to test stability.
To start, we let the motherboard select CPU voltages automatically as we raised the multiplier. We made it up to 4.2GHz without issue, and CPU-Z reported a CPU voltage of 1.2V at that speed. The system blue-screened at as soon as we started our stress test at 4.3GHz, though. Setting the CPU voltage to 1.25V kept the loaded system stable for a couple of minutes before the next BSOD, so we added more. In the end, 4.3GHz required 1.275V.
That voltage bump was enough to sustain our CPU up to 4.5GHz. However, blue screens at 4.6GHz forced us to nudge the CPU up to 1.3V. At 4.7GHz, the chip needed 1.35V, and its core temperatures spiked up to 84°C with regularity—even with the water cooler's fans and pump going at full blast. Thermal throttling didn't rear its head until we tried for 4.8GHz, though. Keeping BSODs at bay at that speed required 1.375V, and the additional voltage sent temperatures into the 90s. No amount of further tweaking produced a stable, throttle-free config at 4.8GHz.
At 4.7GHz, the system was stable enough to run a handful of benchmarks. The x264 test crashed on the first run but completed two subsequent three-loop sessions without issue.
(Although the screenshot above shows a CPU voltage of 1.376V, the CPU was set to 1.35V in the firmware. Asus' AI Suite software agreed with CPU-Z's reading.)
The numbers are pretty close to what one might expect. Pushing the Core i7-4770K to 4.7GHz increases performance by 19-26%, which is in-line with the rise in clock speeds. Our 4.7GHz overclock works out to increases of 21% and 27%, respectively, over the chip's maximum Turbo frequencies for single- and all-core loads.
As is always the case with overclocking, your mileage may vary. That said, it's worth noting that we hit 4.9GHz with a similarly early Ivy Bridge sample a year ago. That CPU also required 1.35V, but it got by with a dual-fan air cooler.
All indications point to overclocked Haswell processors requiring more aggressive cooling than their Ivy predecessors. The Core i7-4770K does have a higher TDP than the 3770K, but the associated heat is also spread over a larger die area. The 4770K's TDP per area works out to 0.47W/mm², while the 3770K's is 0.48W/mm². Haswell and Ivy seem to be on even footing in that regard. The die layouts follow the same basic blueprint, as well.
Haswell and Ivy Bridge also use a similar interface material between their dies and external heat spreaders. Intel used to employ a fluxless solder between those two pieces, but it switched to thermal paste with Ivy.
We don't have a definitive explanation for Haswell's apparent need for most robust cooling, but the chip's integrated VRM may play a role. Voltage regulation was handled off-chip in Ivy Bridge, but Haswell brings it—and the associated heat—onboard the die. Integrated voltage regulation is a big part of Haswell's appeal for mobile platforms. Unfortunately, it may also limit the processor's overclocking potential on the desktop.
130 comments — Last by indeego at 11:29 AM on 08/28/13
|Intel's Core i7-6700K 'Skylake' processor reviewedEnthusiasts get the first taste of a new architecture||486|
|AMD's Carrizo brings power savings to mainstream laptopsExcavator and GCN combine at 15W||83|
|Intel's Broadwell goes broad with new desktop, mobile, server variants14-nm chips for everyone||167|
|The TR Podcast bonus video: AMD, Zen, Fiji, and moreWith special guest David Kanter||54|
|AMD: Zen chips headed to desktops, servers in 2016Details of its new x86 CPU and plans revealed||250|
|Inside ARM's Cortex-A72 microarchitectureThe next-gen CPU core for mobile devices and servers||42|
|Semiconductors from idea to productThe story of how chips are made||56|
|Intel's Xeon D brings Broadwell to cloud, web servicesA big compute node in a small package||40|
|Apple's A9 impresses and the Nexus strikes back: The TR Podcast 188||29|
|Microsoft acquires Havok physics engine from Intel||80|
|AMD unleashes mobile Tonga with the FirePro W7170M||13|
|Deals of the week: Crucial's MX200 500GB SSD and more||10|
|Report: TSMC makes around 6 in 10 Apple A9 SoCs||19|
|Mobile Quadros bring Maxwell to 15" and 17" workstations||4|
|Report: Amazon to halt sales of Chromecast and Apple TV||41|
|The Tech Report Podcast is live on Twitch||2|
|A billion Android devices could be vulnerable to Stagefright 2.0 bug||50|