Optimized Overclocking And Load-Line Calibration
As you overclock to progressively higher clock rates, the only time you increase Vcore is to obtain a more aggressive frequency setting. With this in mind, we revisited our two Core i9-7900Xes to suss out the optimal operating points for temperature and power consumption.
The temperature curves are similar to those obtained for power; one always correlates with the other.
- Up until 4600 MHz, it was relatively easy to get big frequency increases from small Vcore adjustments. That meant we were in a range conducive to overclocking. An 8% overclock across 10 cores led to a rise in power consumption of "only" 22%. In this zone, our frequency gain to consumption ratio is close to 1/3.
- Moving past 4600 MHz, a 5% overclock caused a 23%-higher power consumption, yielding a ratio close to 1/5.
- Beyond 4800 MHz, a 2% overclock led to a 24% rise in power consumption, meaning a ratio of 1/12! At that point, there's little reason to chase even higher frequencies.
The Effects of Load-Line Calibration
When the CPU is idle, using very little energy, our 1.3V setting is easy to obtain. Once a demanding test launches, however, the voltage might drop to 1.26V. An even more taxing workload pushes it lower still. Obviously, that's not what you want to see from a stable system. Fortunately, many motherboards come with a function called Load-Line Calibration. Once activated, LLC artificially raises the supply voltage to compensate for this effect.
In the interest of thoroughness, we ran some tests to quantify the effects of LLC on Vcore. Notice that the measured voltage at idle was always lower than its BIOS setting, though not dramatically so since there was no risk of instability. In Prime95 (without AVX), LLC modes 1 though 5 delivered a Vcore slightly higher than what we asked for, while modes 6 through 8 were a little lower. Prime95 (with AVX) is an even more strenuous test, so modes 1 through 3 over-supplied the Vcore by a substantial amount. At a 1.2V setting, we saw a real voltage of more than 1.23V. Starting with mode 4, the Vcore more closely matched our set point.
Interestingly, we found that the effect of LLC wasn't identical from one sample to the next. Now, we know that Skylake-X-based processors are equipped with an internal voltage supply, so the motherboard doesn't directly supply Vcore. It instead supplies Vccin, which helps explain the small changes we just saw. With that bit of information in-hand, we restarted our tests and kept an eye on Vccin, rather than Vcore.
For a set point of 1.72V, we observed that mode 1 delivers voltages as high as 1.79V (+0.07V) under heavy load. At the opposite end of the spectrum, mode 8 is too weak; its voltage dropped as low as 1.69V (-0.03V). The ideal compromise seemed to be mode 5, which demonstrated balanced results.
MORE: Best CPUs
MORE: All CPUs Content