Last week, I held off reporting estimated power efficiency numbers for AMD's Ryzen 7 family against those of Intel's Core i7-8700K in my Coffee Lake review. The power consumption numbers I observed for the Ryzen 7 CPUs during my tests were unusually (and improbably) high even though performance remained in line with expectations. I didn't track down the precise cause of the issue, but a reinstall of Windows on our AM4 test rig today fixed the problem. With everything in order again, I finally have load power numbers and estimated task energy calculations to share.
Blender is useful for estimating the amount of energy our CPUs consume over time because it's an unusually steady-state workload—power consumption varies little over the course of a render. We use the "bmw27" test file and monitor system power consumption using our trusty Watts Up power meter. Since Blender reliably consumes about the same amount of power over time, we can estimate task energy in kilojoules by relying on the fact that a joule equals one watt expended over one second.
First off, we have instantaneous load power numbers for our test systems. Among our ostensibly 95W CPUs, the Core i7-8700K sips the least system power of the bunch by a small margin. Overclocking the Ryzen 7 1700 and the Core i7-8700K results in a small advantage for the Intel chip, but the emphasis is on small. For all intents and purposes, a system with a Ryzen 7 1700 at 4 GHz draws the same amount of power as a Core i7-8700K overclocked to 4.8 GHz for AVX workloads (like Blender) does.
Instantaneous system power draw doesn't tell us anything about efficiency, though—low power consumption isn't worth much if it takes a CPU eons to render the bmw27 test image. The Core i7-8700K delivers by slotting between the Ryzen 7 1700 and Ryzen 7 1700X, but the Ryzen 7 1800X is the winner among stock-clocked chips.
To visualize the efficiency of each part on the bench, we can plot each chip's estimated task energy against the time it takes for a given CPU to complete our bmw27 test render. The most efficient chips will tend toward the bottom left of the chart, where time-to-completion and task energy are lowest.
As you can tell by the proximity of the dots on our scatter, the Core i7-8700K consumes an estimated 2% less energy than the Ryzen 7 1800X, and about 4% less energy than the Ryzen 7 1700X, but it's anywhere from 2% slower against the 1700X to 7% slower against the 1800X. That's a wash from where I'm sitting. If you're going for the least power consumed for rendering at the expense of a slightly longer task-to-completion, the Ryzen 7 1700 emerges victorious. It's also the only 65W CPU on our test bench, so keep that in mind.
Overclocking the Ryzen 7 1700 and i7-8700K does let them finish the bmw27 render faster, but the extra power consumed this way outstrips the time saved by hastening the pace of number-crunching on both chips. The Ryzen 7 1700 at 4 GHz finishes about 4.4% faster than the i7-8700K at 5 GHz, and it consumes 2% less power. Given that we're estimating these numbers and not measuring the exact amount of power consumed, I'd call that another wash. AMD might have a slight performance-per-watt edge, but we'd need more precise data to say for sure.
Overall, these new results don't change any of the conclusions in my initial review. Both AMD and Intel offer outstanding power efficiency in this generation of CPUs, and our tests back that up. Intel won this battle on other fronts. Only one thing is for sure: the ongoing war for CPU supremacy continues.
|TR's 2017 Christmas giveaway: eight days left and counting||4|
|Rumor: Ryzen 2 set for Q1 2018 and a Fenghuang APU breaks cover||17|
|MSI gives Radeon RX Vega cards an Air Boost||13|
|Corsair's latest SO-DIMM kit takes 32 GB of DDR4 to 4000 MT/s||6|
|Report: Intel Inside co-marketing program will get a budget cut||28|
|Gingerbread House Day Shortbread||17|
|iMac Pro details and release date come into focus||49|
|Radeon Software Adrenalin Edition: an overview||26|
|Tuesday deals: NVMe storage, a GeForce GTX 1080 Ti, and more||9|
|Full disclosure: while I work for Intel; the opinions I express here are my own I think I understanding the issue you ran into. For the Braswell platf...||+35|