Update 10/6/2017 10:45 PM: This article originally stated that the Gigabyte Z370 Aorus Gaming 7 motherboard ships with "multi-core enhancement" enabled. The board in fact ships with the feature disabled. I deeply regret the error and have corrected the article accordingly.
Intel's Core i7-8700K has proved to be an exceptionally well-rounded CPU in our testing so far, but one potential negative has come up again and again in the other reviews I've been reading. Many reviewers have noted that the chip "runs hot," so much so that the idea even made for sub-headline news at one outlet. I was a bit confused reading these statements, because the i7-8700K didn't seem to be an exceptionally hot-running chip in my testing compared to other modern Intel CPUs. Although I ran into a thermal limit while trying to boost voltages enough to get our chip stable under a Prime95 AVX workload, running all of the chip's AVX units at 4.8 GHz was no small feat, and we expect high temperatures as a matter of course from unmodified Intel CPUs when they're overclocked.
Still, the TR water-cooler meeting this morning produced an interesting line in the sand for whether a chip is difficult to cool: can it be held in check by Cooler Master's evergreen Hyper 212 Evo? That $30 tower remains a fine bang-for-the-buck contender among CPU heatsinks, so it's a natural baseline for establishing whether a chip is tough to keep frosty. I don't have a Hyper 212 Evo here, but I do have Cooler Master's MasterAir Pro 4, a 120-mm tower that's basically the same heatsink as a Hyper 212 Evo with a newer fan design. It was simple enough to see whether the Core i7-8700K fell on the right or wrong side of the MasterAir Pro 4's cooling power, so I popped off the 280-mm Corsair H115i that usually cools our test chips and set up the MasterAir Pro 4 in its place.
First off, it's worth defining what "hot" means in the context of the i7-8700K. Intel's Tjunction specification for this chip remains the same 100° C it's been for Skylake and Kaby Lake K-series CPUs. Hit that temperature, and the i7-8700K will begin to throttle. We obviously want to stay as far below that threshold as possible, but it establishes an upper limit for what a "bad" temperature might be for the chip.
With that in mind, I ran the Prime95 Small FFTs torture test at stock speeds to establish a baseline for the chip's thermal behavior. Prime95 hammers a chip's AVX units in a way that's meant to produce the most heat possible, well beyond what any real-world workload might generate. Gigabyte's Easy Tune utility reported that the chip was running at 4.3 GHz—its normal all-core Turbo speed—at 1.1V under this synthetic load. With the MasterAir Pro 4 on top, those clocks and voltages resulted in a CPU package temperature of 78° C, according to HWiNFO64.
Those numbers are certainly warm for a stock-clocked LGA 1151 CPU, but it's worth remembering that we're now asking the cooler to wrangle six cores and 12 threads instead of four cores and eight threads. That's entry-level high-end desktop territory, so slightly higher temps than we're used to should be par for the course. In any case, the stock-clocked i7-8700K proved perfectly happy under our Hyper 212 Evo stand-in.
Next up, I tried to run the chip with Gigabyte's "multi-core enhancement" turned on. This "enhancement" (happily left off by default, as "Auto" means in the Z370 Aorus Gaming 7's firmware) runs all six cores of our i7-8700K at the single-core Turbo Boost speed, or 4.7 GHz. We vigorously search out and disable these kinds of settings for every CPU review we do, since they're the same as overclocking. Other sites may not, and that's not ideal. Not only do these settings ruin any sense of what "stock" performance is from a given processor, they place the same demands on heatsinks as an equivalent overclock would.
I know that's stating the obvious, but we've had bad experiences with these "performance-enhancing" tweaks in the past when they've goosed many-core chips like the Core i7-6950X, and they're sometimes on by default in firmware from Gigabyte and Asus, at least. Readers and YouTube-watchers should be asking whether reviewers explicitly went to the effort to turn off these features before making sweeping conclusions about a chip's power consumption, heat production, performance, and efficiency.
We are glad that Gigabyte's Z370 firmware makes the correct choice with regard to multi-core enhancement behavior, though, and we hope other motherboard brands have followed or will follow suit.
Regardless, I fired up our system in this state and cued up Prime95 Small FFTs again. The chip proceeded to throttle on several cores with a 1.308V Vcore (a difficult figure to monitor given the plunging core clocks, but I tried). That throttling meant the chip was running into its 100° C Tjunction limit on some cores, so the motherboard's automatic voltage control is probably a tad too aggressive given my manual overclocking experience. I also tried running Blender with multi-core enhancement enabled, and while all of the cores got to around 89° C under that load, the chip didn't throttle. That result still suggests a Hyper 212 Evo-class cooler probably isn't sufficient for holding the overclocked i7-8700K in check, given how little headroom it offers.
This behavior shows why "multi-core enhancement" is undesirable: it's overclocking through and through, and it requires cooling to match. Builders who are buying heatsinks under the assumption they'll be facing all-core Turbo speeds of 4.3 GHz from the i7-8700K could be surprised if their motherboard tries to "help" by modifying Intel's factory Turbo Boost behavior. Our Gigabyte Z370 Aorus Gaming 7 test motherboard commendably ships with the feature disabled, but we'd imagine the feature could still catch both reviewers and builders alike off guard. We've been protesting this "feature" for years, and we'll continue to do so when it rears its head.
Finally, I tried the same manual overclock I achieved with our Corsair H115i liquid cooler: 5GHz with a -2 AVX offset and a dynamic Vcore in a range of 1.284V to 1.296V. Under the MasterAir Pro 4, running Prime95 caused the chip to throttle, while Blender caused it to run in the low 90° C range. Considering that my overclock was pulling another 100 MHz from the chip's AVX units with only slightly less voltage, it's not a surprise that I got similar thermal results. Under these conditions, the chip definitely exceeds the informal "difficult to cool" barrier that we drew at the beginning of this article.
For comparison, Corsair's 280-mm H115i produced a 90° C package temperature and core temperatures ranging from about 84° C to 90° C using the same settings and voltages with Prime95 Small FFTs. Blender topped out our overclocked i7-8700K at about 80° C at the package. The H115i definitely reins in the i7-8700K if you're shooting for the ability to run Prime95 for hours, as one might want to do for extreme stability testing.
These are all rough benchmarks, but at the end of the day, Coffee Lake does seem to run hotter at stock speeds than the quad-core CPUs that have come before it. That's probably as it should be: there are two more cores and four more threads to deal with under the heat spreader. Builders planning to cool the chip at stock speeds should certainly be able to get away with an inexpensive cooler like a Hyper 212 Evo, but those hoping for a Prime95-stable overclock without a delid and repaste need to budget for a substantial liquid cooler. In that sense, the i7-8700K is no different than the Core i7-6700K and Core i7-7700K before it, and it's definitely harder to cool than AMD's Ryzen CPUs. AMD's chips all boast soldered heat spreaders, and metal is undeniably a better thermal transfer medium than paste.
The question of a paste-based TIM versus solder is almost certainly the largest variable in keeping Coffee Lake on ice relative to Ryzen CPUs, but I think there's more to it than that. First off, it's worth noting that Intel's implementation of AVX in the Skylake microarchitecture offers two 256-bit vector units per core, while the Zen architecture only offers two 128-bit-wide units per core. Skylake also has wider data paths that need more wires to implement, and that presumably means higher power usage when moving data around. When we run an intense AVX workload like Prime95, then, the stress test should unsurprisingly do more work, consume more power, and ultimately generate more heat on a chip that's capable of sustaining twice the SIMD throughput. It's certainly easier to cool an overclocked Ryzen CPU thanks to its soldered heat spreader, but it's hard to argue that one isn't getting more out of overclocking the Core i7-8700K in many tasks despite its higher temperatures. That fact should be part of the value consideration when setting out to overclock either chip.
Whether Intel is doing the best it can to support overclocking on its chips through its thermal interface material of choice is another question, and it's one that's raged since Ivy Bridge and coursed through Devil's Canyon. Coffee Lake doesn't do anything to quench the flames. Folks seeking the lowest load temperatures and highest possible overclocking headroom from Coffee Lake chips will likely need to reach for liquid-metal TIM, their delid tool of choice, and a hefty liquid cooler or giant tower heatsink. At stock speeds, though, the i7-8700K should be fine with the same Cooler Master Hyper 212 Evo that's graced countless systems. Just be sure to terminate any multi-core enhancement settings in your motherboard's firmware with extreme prejudice first.
|Qualcomm shows progress on 5G mobile broadband||14|
|ROG Strix X370-I and B350-I are itty-bitty boards for Ryzen builds||10|
|Samsung foundry train stops at 8-nm LPP before heading to EUV||12|
|Wednesday deals: a Ryzen combo, mechanical keyboards, and storage||9|
|RX Vega prices inch downward in our latest graphics-card spot check||25|
|HP ZBook x2 detachable is a consummate professional||7|
|NZXT Grid+ v3 keeps PCs quiet with machine learning||9|
|Razer's Blade Stealth and Core V2 step to the cutting edge||14|
|Intel unveils purpose-built Neural Network Processor for deep learning||19|