Aurich Lawson / Thinkstock

Future of mobile CPUs, part 2: What’s ahead for the major players?

Apple and Samsung face different challenges than Nvidia, Intel, and Qualcomm.

by David Kanter – Feb 22 2013, 11:50am EST

In the first part of our series, we explored the major trends that will influence the mobile system-on-a-chip (SoC ) market over the next five to ten years. This sets the backdrop for looking at the architecture for future SoCs and the specific players within this market, both critical IP players as well as the actual SoC vendors. For the most part, this focuses on mid-range to high-end devices, rather than the lowest-end smartphones and tablets. This means that some SoC vendors have been omitted, for the sake of clarity and brevity.

SoCs today

The vast majority of smartphones today are single- and dual-core SoCs. At the very high-end, there is a smattering of quad-cores. The same is mostly true of tablets, although the larger power budget means that the processors tend to skew towards higher core counts. The CPU cores are clocked at around 1GHz, and the more advanced ones feature out-of-order execution and modest superscalar issue, typically two to three RISC instructions per cycle at peak. Simpler cores for more power-constrained systems tend to be in-order and issue one to two instructions per cycle. This level of complexity is generally on-par with the CPU cores found in the early to mid 1990s.

Realistically, it is hard to see any benefits from quad-cores in mobile devices. The majority of PCs today sell with dual-core CPUs, and that is a reflection of the state of software; multithreading is hard and most applications are single threaded. Software for mobile devices is even more primitive and less amenable to threading. Comparing a quad-core to a dual-core at the same power, the dual-core should be able to reach about 25 percent higher frequencies (power scales roughly with frequency cubed). For the vast majority of workloads, a faster dual-core CPU will have better performance. Despite this fact, there appears to be some marketing value for quad-core SoCs, even if the delivered value is minimal.

One reflection of the divergence of smartphones and tablets is the graphics for these devices. Tablets have higher-performance graphics to drive the larger and higher-resolution displays and to make use of the greater power envelope. The actual GPU cores are usually the same, but with more cores and higher frequency for tablets. Looking at the iPhone 5 and iPad 4, the latter GPU has about 3X the shader throughput measured in FLOP/s (~100 vs. ~30 GFLOP/s). In terms of performance, the iPhone 5 is roughly the equivalent of a very low-end discrete DX10 GPU from 2007, while the iPad 4 resembles a mid-range model.

The other significant blocks in a mobile SoC are the wireless modem, which is often discrete for high-end phones and tablets (i.e., LTE devices), along with dedicated hardware for video encode/decode and image processing for the camera.

Power management ties together all these blocks and is particularly vital, since performance is limited by both the battery life and skin temperature (i.e., how hot the case gets). Simply put, there isn’t enough power or cooling for every block to be in a high-performance mode simultaneously. For example, when running a strenuous game, the display and GPU will draw much of the power; the CPU will actually have to reduce frequency and voltage to deliver the best overall performance. This becomes even more complex if there is significant wireless traffic as well.

SoCs of the future

Looking out 5-10 years, Moore’s Law means that transistors will be even cheaper. However, battery technology improves slowly and the maximum skin temperature is constant. Consequently, power will be even more of limiting factor in the future than it is today. So techniques that spend transistors (or area) to reduce power will be increasingly attractive.

While change is slow, eventually mobile developers will be able to take advantage of multiple cores. At this point, quad-cores can be more efficient by reducing frequency and voltage, as the PC industry has shown. Most workloads will still be single threaded and need high frequencies, so the SoC must be able to efficiently deliver both aggregate throughput and single-core performance. Eventually, almost all mobile SoCs will move to quad-core to handle the few cases of properly parallelized code.

Read more: Future of mobile CPUs, part 2: What’s ahead for the major players? | Ars Technica.

Home           Top of page