That eight-core smartphone isn’t as fast as you think it is

That eight-core smartphone isn’t as fast as you think it is

Ever since the first quad-core smartphone and tablet designs began to show up in market, we’ve been asking whether or not these devices offered meaningful performance improvements over lower core counts. Not only are there software issues to contend with, since without multi-threading support, adding new cores is literally useless, there are also significant concerns around power consumption. Adding more cores means more heat, and smartphones, with ~2.5W operating envelopes, just aren’t designed for high-speed multi-core operation.

A new paper from Moor Insights & Strategy delves into these questions and doesn’t come away with many wins for the “Add more cores” team. Extensive testing shows that the gains associated with adding CPU cores taper sharply above four, while performance sometimes regresses moving from four cores to eight. The analysts compared multiple devices, including the LG G4 (Snapdragon 808, 2x Cortex-A57 + 4x Cortex-A53), the Xiaomi Mi 4i (8x Cortex-A53), LG G Flex 2 (quad-core 4x Cortex-A57 + 4x Cortex-A53) and the HTC Desire 820S (8x Cortex-A53). The full paper is available for download, but here are two of the highlighted workloads.

Analyst Anshel Sag used third-party software to disable certain cores on the devices in order to benchmark the impact of running fewer cores in common tests. In PCMark’s Android test, the eight-core Cortex-A53-based Xiaomi Mi 4i shows no scaling between two cores and eight cores, while the LG G Flex 2’s performance holds steady from eight cores to four, before declining sharply at two cores. The team opted to use PCMark because it explicitly attempts to create a mobile phone-specific workload test rather than running a purely synthetic set of benchmarks like Geekbench or AnTuTu.

In most cases, customers would likely shrug if told that having more cores doesn’t always help, assuming that the performance will be beneficial in certain workloads. The team’s analysis of the Basemark X benchmark, however, shows that’s not always the case. Check out what happens to the LG G Flex 2 when core counts are reduced:

As cores shut off, overall performance rises, particularly at the dual-core mode. With most of the chip offline, performance jumps. One potential conclusion that would explain these results is that applications are poorly threaded, which prevents them from taking advantage of higher core counts. But the fact that performance increases at the two core mark suggests something more fundamental at work — the chips in question are hitting their thermal trip points unless more cores are shut down.

Physics is not a friend to marketing

All of this is inconvenient to manufacturers who are pushing eight-and-ten-core devices because the public has picked up on core counts as an easy marketing tool for selling smartphone chips, but the conclusions are inescapable. No big.Little configuration or implementation of DVFS is going to fix multi-core smartphones. With voltage scaling improving only slightly, the power consumption improvements that manufacturers have been able to offer at smaller nodes have shrunk. Once upon a time, nodes offered 40-50% reductions whereas today, 25-35% is more likely.

That shrink, however, has not been reflected in smartphone design. There’s an innate power consumption hit you take for having silicon turned on, even if you’re operating cores in a low-power standby mode. Companies like Qualcomm, MediaTek, and Samsung have also wanted to push the envelope on clock speed scaling as well, and typically push to deliver slightly higher envelopes (or, in MediaTek’s case, expand the chip count even further). Add more cores and a few hundred more MHz (along with a higher-resolution display) and you’ve neatly soaked up all of the power savings that the foundry was able to deliver.

In theory, an SoC and software stack that could perfectly anticipate power demands could take advantage of more cores by expertly power gating CPUs that weren’t necessary and only ramping up enough cores to deliver maximum performance. An application that mimicked BaseMark X, in this case, would haul back the CPU to just two cores, in order to give the GPU more frequency headroom. In practice, that level of intelligence simply isn’t available. Applications typically aren’t tuned for each and every SoC (the number of SoCs in-market make this impossible) and it’s not clear if Android even allows for the ultra-fine-grained clock control that would be required in any case.

Applications like Geekbench have become popular as a way to demonstrate the theoretical performance of smartphone SoCs, but real-world application testing shines a different light on things (Moor Insights also tested camera and chat applications to round out their multi-core evaluation). As things stand, there are some benefits to quad-core devices and virtually no gains from octa-core. In a few cases, moving to more cores actually makes things worse. The overall situation will depend on which applications you use, but the relentless push to add cores is doing end-users no favors.