The biggest takeaway from our livestream was the determination that the i7-8086K, much like Ryzen 2000 CPUs, is able to sustain significantly lower voltages at a given frequency than its 8700K predecessor. Again, the CPUs are derived from the same architecture – even the same wafers – but one is pre-tested and sorted (“binned”) for high-frequency 8086K CPUs.
Blender 8086K CPU Binning
We started our bin testing just with Blender and kept things simple. Blender is significantly less brutal than Prime95, although we have done testing with P95 8K FFTs. With Prime, the trick is that smaller FFTs generate more heat and require more power – but 8K FFTs, while realistic to some scenarios, are a bit too abusive for our quick testing. We wanted to validate for Blender (production stand-in) and Firestrike (gaming stand-in).
In Blender, we found 5.1GHz at 1.35V to be stable, with other results above.
Firestrike 8086K CPU Binning
Firestrike is a lot less intensive than AVX workloads like Blender or Prime, but more important for our competitive benchmarking purposes.
5.0GHz at 1.30V was sustainable during our overclocking livestream, which we later verified as stable for prolonged periods of Firestrike Physics testing. This is functionally a stand-in for gaming, and illustrates just how well the binned 8700K CPU performs. Our 8086K holds 5.0GHz at 1.30V to 1.35V for Firestrike, whereas our 8700Ks hold 5.0GHz at 1.40 to 1.45V, depending on which CPU it is. That’s the bigger takeaway – at a given clock, we’ve reduced our voltage requirement significantly.
5.1GHz also held at 1.30, but only for long enough to run a 30-second test. It did not survive longer burn-ins.
We next found a 53x multiplier to be stable at about 1.39V, which is also impressive when considering most 8700Ks could never reach 5.3GHz without exotic cooling. That’s even more true now, since most of those higher bin 8700Ks have been removed to become 8086Ks.
We were unable to stabilize a 54 multiplier under any circumstances, including per-core binning. We were, however, able to sustain a 101 BCLK, shown in the left-most column, with a 53x multiplier and 1.41V. This gave us a 5350MHz core clock at 1.41V.
For purposes of Firestrike – Physics, it would appear that the optimal performance-to-thermal and power configuration would be roughly 5.1GHz at 1.32-1.35V for our CPU. As always, this will change based upon each individual CPU.
The ASUS Maximus X board seems to keep a fairly flat voltage provision for Vcore when at LLC level 6, which is what we want. Going to LLC level 7 was sometimes necessary, and tended to overprovide voltage when spurious delivery became problematic.
Firestrike CPU Physics Scores – 8086K
As for relative scores and scaling, we found the following results across our 8086K frequency testing.
Firestrike spat-out an 18007 score for the complete baseline, stock score with just XMP enabled. Clocking trivially to 5.0GHz at 1.3V got us to 20900 points, an increase of 16% from about a 700MHz bump in clock.
We next observed only a 1-2% increase in score by overclocking an additional 100MHz, up to 5.1GHz. Our 5.2GHz overclock at 1.35V yielded an increase to 21773 points, an increase of 21% over the baseline 18000 score from earlier. At 5.3GHz and 1.4V, we observed an increase to 21992 points, another marginal hike over the previous hike to 5.2GHz. Each 100MHz increase isn’t individually gaining a lot, but it is measurable and repeatable.
By jumping to a BCLK of 101, which increased memory frequencies, we ended up at 23223 points – a big jump, but a lot of that is from the BCLK increase. Going to BCLK 101 with 3600MHz memory at CL15, including an uncore of 51-52, we had a final result of 23745.
Again, the real takeaway with the 8086K is sustained high frequencies with low voltages. When high clocks can be run on 1.2-1.3V, that’s a major power and thermal win, and translates to a major acoustic win. Overclocking the 8086K is the same procedure as for the 8700K, it’s just easier – based on samples already tested – to hit those higher, 5.0+GHz frequencies.
Editorial, Testing: Steve Burke
Video: Andrew Coleman