About two weeks ago, we covered a report from overclocker der8auer, who claimed that the X299 VRM cooling situation was a disaster due to improper cooling decisions from a number of motherboard manufacturers. This kind of claim is always controversial, but it’s now been verified — and to some extent, sharpened and clarified — by more advanced testing.
Tom’s Hardware has published an extensive article on how the top-end Skylake-X parts, like the Core i9-7900X behave under heavy load in a variety of conditions and when mounted with top-end water cooler chillers that can run you $ 1,500 per system. But the problem isn’t limited to the VRM cooling issues that can hit X299 motherboards, though Der8auer wasn’t wrong about that issue. Intel’s decision to use thermal paste for a 10-core Skylake-X CPU with a turbo frequency as high as 4.3GHz is causing a huge problem all on its own.
What this graph shows is the CPU temperature inside the die (as reported by the CPU itself), the temperature of the water inside the high-end Alphacool Eiszeit 2000 Chiller + Alphacool Eisblock XPX, and the temperature of the CPU heat spreader. With the CPU core at idle, there’s just an ~9C difference between the heatsink temperature and the actual internal CPU core temperature. But as the CPU heats up, that gap explodes. When the CPU core hits 100C, the heat spreader is still just barely 30C. In other words, increasing the internal temperature of the CPU by 3.33x increased the temperature of the heat spreader by 1.38x. And that’s with a $ 1,500+ cooling system. Delidding CPUs isn’t trivial, particularly with the integrated VRM that Skylake-X uses, but the thermal paste on Skylake-X is going to kill its scaling.
VRMs and system heat
Again, I recommend reading the full story, but the bottom line is this: Even at 160W, Skylake-X can’t run a high-end air cooler like the Noctua DH-15 for more than a few minutes before it begins to throttle. At least a decent CLLC (closed loop liquid cooler) is required, and a top-end system is mandatory for full performance. These settings, however, do not push temperatures on the VRMs or other components all that high. Keep in mind, no overclocking is being done in either of these cases.
Step up to a maximum TDP of 230W, however, and that changes. VRM temperatures hit 90C. This next observation from THG is worth quoting in full:
Observation #4: Depending on whether you use the Enhanced Turbo option or not, power consumption in excess of 230W (in AVX-heavy workloads) or as high as 200W (without AVX) is possible right out of the box. That’s no longer in the realm of air cooling. You need a good closed-loop liquid cooler at minimum. Even that’s going to hit its limit, though, since the core temperatures peak at close to the point where thermal throttling begins.
I should note Noctua has represented to me that its best coolers can cool a 250W CPU, but that may be a chip that doesn’t have the VRM overheat and internal CPU temperature problems than Skylake-X does.
In the end, Skylake-X clearly has a heat problem. Even with a $ 1,500 chiller and a water temperature below 30C, the THG team couldn’t overclock to 250W – 300W and keep performance consistent, no matter what coolers they used. High core-count chips have intrinsic overclocking issues related to the sheer number of cores they pack into a wafer, and THG notes that the power curves on Skylake-X resemble a high-end GPU; worse even than AMD’s FX-9590.
I wasn’t planning to put a major focus on overclocking for Core i9, but I’m increasingly convinced this aspect of performance needs to be examined. Skylake-X clearly has issues, and while Intel can fix a lot of them by just using solder instead of paste, that’s not going to comfort anyone who bought the platform thinking they were getting some great overclocking chip.