Thermals and Overclocking
The Haswell and the LGA1150 platform change the overclocking procedure for two reasons. There are new divisors for the PCIe/DMI bus and there is now a voltage regulator right inside the CPU.
The first thing was expected because the fixed correlation between the base clock rate and the PCIe/DMI clock rate used to make it impossible to overclock CPUs by raising the former. The PCIe/DMI bus doesn’t work well when its frequency deviates from the default 100 MHz, so increasing the base clock rate to 105-107 MHz used to render the LGA1155 platform nonfunctional.
This problem is solved to some extent in the new Haswell processors. There are now a few dividers that let you set the base and PCIe/DMI clock rates not only as 1:1 but also as 5:4 and 5:3. Thus, the LGA1150 platform can be stable at a base clock rate of 100, 125 and 166 MHz. All crucial internal frequencies remain at their defaults in this case, but the x86 cores, the uncore part, the integrated graphics core and system memory get overclocked proportionally. It also means that even LGA1150 CPUs with locked multiplier can be overclocked, but only by 25% or 66% above the default frequency.
Only K series CPUs with an unlocked frequency multiplier offer you complete overclocking freedom. By the way, during this the transition to Haswell Intel added yet another bit to the processor multiplier register, so the maximum multiplier setting during overclocking has now reached 80x.
The integration of the voltage regulator into the CPU affects overclocking, too. The integrated regulator behaves in a peculiar way. The CPU voltage used to drop at high loads, but the integrated regulator, on the contrary, automatically increases it in this case – by over 0.1 volts even at default settings. This effect is more conspicuous at overclocking.
Unfortunately, this behavior is typical of the Haswell. You can only eliminate it by locking the voltage at a certain level, which disables all power-saving technologies. So, Haswell offers a difficult choice: you either have to put up with high temperature and heat dissipation of the CPU under heavy loads caused by automatic increase in the processor core voltage or to give up power savings in idle mode.
But this is just part of the problem. Haswell turned out to be much hotter in real life than its predecessor. The maximum permissible temperature of its CPU cores is 100°C but even in nominal operational modes Core i7-4770K would get as hot as 75-80°C even with a high-performance air-cooler.
To illustrate Haswell’s thermal performance we performed a quick comparison between Core i7-4770K and Core i7-3770K working in their nominal mode and tested with the same NZXT Havik 140 cooler:
The Haswell CPU core temperatures are seriously higher than those of the previous generation processors. And although most every-day tasks do not cause the CPU to heat up so dramatically, we should base our conclusions primarily on specialized stability tests, which create heavy but nevertheless quite realistic load.
So, it turns out that overclocking the new CPUs calls for much better coolers than those we could use for Ivy Bridge processors. In other words, it is harder to reach the same results when overclocking Core i7-4770K as we did with the overclocker-friendly Sandy Bridge and Ivy Bridge products in LGA1155 form-factor.
For example, we only made our Core i7-4770K CPU stable at 4.4 GHz. The temperature of the CPU cores was alarmingly high while running the stability-testing LinX-AVX utility, even though we used a very good cooler - NZXT Havik 140.
To achieve the result shown in the screenshot, we had to increase the CPU voltage to 1.2 volts. This was only 0.14 V higher than the nominal Vcore for our particular processor, but nevertheless, the temperatures were through the roof.
Thus, even a small increase in voltage leads to a dramatic increase in temperature of the computing cores, which means that the Haswell microarchitecture is energy efficient at low clock rates and low voltages only. Haswell-based desktop and overclocker-friendly CPUs are not energy-efficient at all. As a result, Haswell’s overall overclocking potential doesn’t inspire much optimism at all. In other words, another iteration in Intel’s microarchitecture is in fact a step back in terms of frequency potential, even though Intel did everything possible to compensate for it by adding extra overclocking-friendly features into their new processors.