Now that Haswell-E’s been with us for about a month, we’ve had a chance to study it a bit more thoroughly and collect some data. One of the standout aspects of the architecture, particularly in its premium octal-core configuration, is its power consumption. At stock speeds, Haswell-E is an incredibly efficient architecture. However, while mainstream Haswell and Devil’s Canyon are able to see substantial overclocks without massive corresponding increases in power draw, Haswell-E can quickly lose that efficiency edge when overclocked.
We used a sample of four Intel Core i7-5960X CPUs and tested overclocking on each one to get a fairly holistic idea of what we can get. Unfortunately the results weren’t as diverse as I’d hoped, likely owing at least partially to these samples all being from the same batch. Power consumption was measured through Corsair Link using the AX1200i power supply.
Our test platform:
- CPU Cooler: Corsair Hydro Series H100i (four fans in push-pull)
- Motherboard: ASUS X99-Deluxe
- Memory: 4x4GB Corsair Dominator Platinum DDR4-3200 CAS 16
- Video Card: eVGA GeForce GTX 780 Ti
- Storage: 128GB and 256GB Force LX SSDs
- Enclosure: Corsair Graphite Series 760T
- Power Supply: Corsair AX1200i PSU
For peak load numbers, the CPU was kept under a sustained stress test for ten minutes.
To start, we’ll look at the voltages required to get each of our samples up to the individual overclocks. We tested them at their stock configuration, then at 3.5GHz all the way up to 4.4GHz. Note that none of these chips was able to do 4.5GHz at under 1.4V, which in this writer’s opinion is just too high for this architecture and manufacturing process.
You can see there wasn’t a lot of variance between these four chips, with the exception of the last one. The fourth chip was able to hit the overclocked speeds at slightly lower voltages than its kin, but still had the same trouble getting to 4.5GHz.
Voltage has a pretty linear relationship to power consumption on these chips, too.
You can see load consumption hangs out at around to slightly above 200W at stock speeds, and considering that’s driving eight cores that’s really not too bad. Manually setting the clock speed to 3.5GHz across all eight cores and the voltage at a steady 0.975V shaves off a little consumption for most of the chips. Go to the other end of the scale, though, and suddenly you’re looking at more than 350W of load power.
When overclocking CPUs, there’s always an “inflection point.” Up to that point, getting an extra speed bin requires a modest increase of voltage if any. On all of these chips, the inflection point has been between 4.2GHz and 4.3GHz. There’s the steady, gradual increase of consumption up until about 4.1GHz, and then suddenly we’re cresting 300W under load at 4.2GHz. 4.4GHz, while totally doable, can add as much as 60W of additional power consumption over 4.3GHz for just a single bin.
This graph should give a clearer idea. Since all four of our samples behaved fairly similar, I’ve isolated one of them and tracked the voltage and wattage scaling against each other. The relationship is fairly linear, with both of them spiking at 4.4GHz. Interestingly, though, voltage and wattage climbing doesn’t start to get really onerous until about 3.9GHz, and this is true across all four chips. While I desperately wanted 4GHz to be the sweet spot, it seems like 3.9GHz is really where it’s at.
When we take a look at the idle to load delta, we get a better idea of just how nasty power consumption becomes at high load. At 3.5GHz, the chip idles at ~104W and under load goes up to ~206W, netting about 102W of power consumption. That’s not so bad. On the far flung, nastiest edge (4.4GHz), though, you’ve got ~138W of idle power – still not unreasonable – but a staggering ~358W of load. Suddenly the chip is adding 220W to your power consumption under load.
Also, like before, note that power consumption starts to really climb after 3.9GHz; up to that point, you’re getting a healthy amount of performance at fairly reasonable power.
You can see from the data here that the i7-5960X is at least reasonable with power consumption at stock speeds, but starts to really beat on your power supply once you overclock past a certain point. For our samples, that point was about 3.9GHz, which is a little disappointing since it’s shy of the magic 4GHz mark. Is a nice, round 4GHz worth about 20W?
We’ll take a look at how much performance overclocking gets you as well as performance-per-watt metrics in a future article, so stay tuned.