Comparing the power efficiency of mobile CPUs

What’s more power efficient? Intel’s Atom, Qualcomm’s Krait, Nvidia’s Tegra, or Samsung’s Exynos 5 Dual?

All four SoCs have very different architectures and performance characteristics, and they’re all built on different processes too, making the analysis rather difficult. Fortunately, Anandtech has collaborated with one of Intel’s power engineers to work it out – and the results might surprise you.

Measuring the power draw of just the CPU and GPU is, as you can imagine, rather difficult. It’s one thing to measure the power draw of the entire device (just measure the draw on the battery circuit), but another thing entirely to measure just the CPU and GPU.

First you have to track down the power circuitry for the SoC, and then you have to find the exact power lines that feed the CPU and GPU on the SoC. Fortunately, Anand had an Intel engineer on hand to find the power lines – and the engineer also came bundled with the £2,000 piece of kit from National Instruments (NI USB-6289) that was necessary to accurately measure the fluctuations in power draw (pictured above).

The following SoCs were tested: Intel’s Atom Z2760 (32nm, Clover Trail, found in the Acer W510); Nvidia’s Tegra 3 (TSMC 40nm, found in Microsoft’s Surface RT); Qualcomm’s Krait APQ8060A (TSMC 28nm, found in the Dell XPS 10); and Samsung’s Cortex-A15 Exynos 5 (32nm, found in the Google Nexus 10).

The following image shows Krait vs. Clover Trail vs. Tegra 3 idle CPU power consumption:

The results are intriguing. For a start, the 32nm Atom outperforms Tegra 3, both in terms of performance and power consumption. Compared to Krait (which is a much newer architecture than both Tegra 3 and Clover Trail), Atom is faster, and is competitive power-wise. Anand notes that Krait’s idle power is generally lower than Clover Trail, but that’s mostly down to Atom’s older architecture and larger process. Before 2013 is through, Intel will finally launch a new out-of-order Atom core built on a 22nm process, and in all likelihood it will blow the competition away.

And then there’s the 32nm Exynos 5, based on ARM’s brand new Cortex-A15. From the get-go we’ve known that Cortex-A15 would be a monster – and indeed, it’s the first ARM chip to comfortably outperform an x86 design from Intel – but at what cost?

Anand’s testing shows that the Cortex-A15 chews through power at an impressive rate, comfortably consuming more than 4 Watts during load – and that’s with Samsung throttling the CPU and GPU when it hits 4 Watts; the actual TDP allows for up to 8 Watts. Compared to Clover Trail, Krait, and Tegra 3, which all draw less than 2 Watts during load, you can see how Cortex-A15 might have some issues finding its way into smartphones.

The following image shows Nexus 10 (Exynos 5, Cortex-A15) power consumption:

Cortex-A15′s excessive power draw also offers an explanation for Krait and Apple’s Swift (A6/A6X) CPU core: Qualcomm and Apple probably took one look at the A15 white paper and decided to design their own, more efficient cores. It will be interesting to see the power characteristics of Tegra 4, a quad-core Cortex-A15 design.

The most interesting takeaway from Anand’s testing, though, is that it actually draws Haswell into the mix. Intel has already shown off third-generation Core (Ivy Bridge) parts with an 8 Watt TDP, and there are rumours that Intel will launch some sub-10 Watt Ivy Bridge parts at CES this week. Haswell (fourth-generation Core) will probably launch at around 8 Watts, but it wouldn’t be surprising if Intel reaches 4 Watts – if not at 22nm, then at 14nm in 2014. Suffice it to say, if Intel can squeeze desktop-like performance into a 4 Watt part, then the future of mobile computing is very exciting indeed.

Image Credit: Anandtech