Benchmarks measure how quickly a given system can complete a given task and are often written to use all available relevant resources. A mobile benchmark might use all eight cores in a smartphone with a traditional big.LITTLE processor core configuration, that is, four lower powered, high efficiency cores paired up with a four higher performance core processor. This only tells us how well the device can complete the benchmark but does not correlate into real world performance or smoothness: many applications only use one or two processor cores rather than all eight. Furthermore, modern System-on-Chips are able to produce significantly more heat than the device chassis can cope with and so are simply unable to deliver their maximum performance for more than a few minutes. Yes; the Qualcomm Snapdragon 810 has received much criticism in the press about the high temperatures the processor (and its host device) reaches, but it’s an issue that impacts all modern processors and devices to a greater or lesser extent. Yes; there are weaknesses in how well the Snapdragon 810 handles its heat, but can Qualcomm really be blamed when a manufacturer shoehorns a processor able to pump out 12 watts of heat into a chassis that cannot dissipate so much? Samsung’s Galaxy S6 and Exynos 7420 also has the same issue too, even though the 7420, built on a smaller manufacturing process of 14nm rather than 20nm for the Snapdragon 810, does produce less heat.
What is the solution to the problem? The answer is to refine SoCs to reduce power and heat output whilst maintaining performance through software changes. This is because newer versions of software tends to be heavier on the device, so in order to stay still the SoC needs to be more powerful. There are a few ways that this may be done, from reducing the process size and so requiring less voltage to drive the circuits. Power consumption (and heat output) is proportional to the square of the voltage applied so there are big gains to be made here. The other approach is to redesign processor cores and accompanying circuitry to improve efficiency. Processor manufacturers are doing both.
Let’s start with the ARM reference cores. We’ve written about the ARM Cortex-A72 core before, but ARM have designed this core to be more efficient compared with the current Cortex-A57 cores that are used in most current 64-bit ARM processors. By the numbers, ARM redesigned the A72 to have a 60% higher instruction-per-clock ratio than the previous generation A57. We’ve also seen power consumption reduction figures of a claimed 20% reducing when the A72 core is built on the same process size as the A57, although this may be the maximum heat output rather than across a more balanced load. Nevertheless, the A72’s improved efficiency should result in the processor running at a lower clock speed, which in turn means it will require a lower voltage and this could significantly reduce heat output for a given (realistic) workload.
We also know Qualcomm are working on their custom Kyro processor core, although details about the makeup of the Kyro are sketchy. We know it’s going to form the basis of the Snapdragon 820 processor and this appears to be a quad-core processor, rather than the Snapdragon 810’s eight cores spread over a big.LITTLE configuration. As such, it is no surprise that multi-core benchmarks show the Snapdragon 820 more or less in line with the older processor – it has half the number of cores, so it will likely underperform in the unrealistic “Chuck Norris” all-cores-blazing tests. Qualcomm, we are sure, will be working on reducing power consumption and so heat, but quite how this is to be achieved remains to be seen. We’ve seen rumors of the Snapdragon 820 being clocked at 3 GHz and the processor will be manufactured by Samsung, the current world leader when it comes to mobile SoC manufacturing.
The next few months are going to be interesting. Qualcomm have already detailed how their mid-range System-on-Chips will be based around the A72, which implies that Kyro will be “better,” however this is measured. ARM’s big.LITTLE designs were all about reducing power and heat consumption because the higher powered A57 cores could produce too much heat for a small, thin chassis. Perhaps Qualcomm’s breakthrough with the Kyro core-based processors is that they combine high efficiency with high performance and the minimum number of cores? NVIDIA’s custom 64-bit core, the Denver, residing in the Tegra K1-64 (in service in the Google Nexus 9) already manages this and of course we can take a look at how Apple use a high performance, 64-bit dual core processor in their mobile products, too. And to jump back to benchmarks, multi-core tests are interesting from an academic perspective, but not so relevant for day to day. A more balanced benchmark for comparison purposes may be something like PCMark.