The battle for the chips that power modern tablets and smartphones has taken an interesting turn in recent weeks. ARM is the undisputed leader in this space, as its low-power chips run in the vast majority of mobile devices, but rival Intel firmly believes it has better technology in the shape of the upcoming Silvermont Atom processors and associated system-on-chip (SoC) implementations.
Intel the underdog
Intel is in the unaccustomed position of playing catch-up in the sub-5W processor space, but the chip giant has been working hard to provide technology that's a viable alternative to the ARM ecosystem. This fight, whose denouement is some way off, strikes at the very core of processor design and has wide-ranging ramifications. Intel champions the x86 instruction-set architecture (ISA) while ARM, whose heritage is very much in mobile-chip design, runs off a simpler RISC-based ISA: the two are incompatible and require different implementations of operating-system software.
ARM's fastest in-production chip design is the Cortex-A15, found in devices such as the Nexus 10 tablet, select Samsung Galaxy S4 smartphones, and Nvidia's Shield handheld console. Intel is targeting this Cortex-A15 for special attention, and Dadi Perlmutter, general manager of the Intel Architecture Group, has been on record as saying "[the ARM] Cortex-A15 is not even close to [Intel] Silvermont. They are higher power and much behind us on performance which means they are on the wrong scale." Bold words indeed for a company who has yet to make a substantial impression in the mobile market and who, crucially, hasn't shipped the 'ARM-beating' product in question.
ARM, on the other hand, fully believes that it continues to have the upper hand in the industry that its designs dominate. Noel Hurley, ARM's vice president of marketing and strategy, beats down the Intel bombast by stating that "Our analysis shows that we're more than a generation ahead [of Intel Silvermont]. We've maintained out leadership in this place."
Who has the better performance? Only time can tell
So who is right and who is wrong? The answer to this riddle is, much like the problem, multi-faceted in nature. Designing very low-power chips for the embedded market and entry-level (smart)phones requires fitting an adequate amount of performance into the smallest possible space - or a high mW/mm² ratio, if you will. Low end-user prices dictate that a mm shaved off chip silicon can be worth more to a low-cost handset partner than increased performance: absolute cost is king. ARM typically provides the Cortex-A5 and Cortex-A7 intellectual property (IP) to its range of partners for this purpose. The upper end of the market, meanwhile, has a different concern, one of fitting as much computational power without breaching an SoC thermal limit of around 4W, including graphics.
Putting it another way, the challenge at the high-end is on providing maximum performance-per-watt figures, with chip size a secondary concern. ARM aims for maximum sustained performance at an energy cost of roughly 700mW per processor core, excluding the power required by any large caches or interconnects. A quad-core Cortex-A15, then, fabbed on a 28nm process, runs deliberately close to the acceptable thermal limit for inclusion in smartphones and tablets, and is significantly thirstier than a same-clocked Cortex-A7. The need for extra energy is down to a number of factors, including a wider design, longer pipeline and the requirement for an out-of-order execution engine on high-performance chips. Crunch the numbers and that mW/mm² ratio actually goes down as chip performance goes up.
It is possible to calculate a Cortex-A15's performance along a throughput-vs-energy curve because it is a shipping product. Intel, on the other hand, hasn't disclosed this same curve for Silvermont... and that leaves an awful lot of room for conjecture. The most up-to-date information states that the upcoming Silvermont chip can consume 5x less idle power and produce 3x the maximum performance when compared to the incumbent Saltwell-based Atom processors shipping today as part of the Clover Trail+ platform.
A nebulous world
And even if you can measure the energy characteristics of Intel's Silvermont to the same degree as Cortex-A15, how do you define performance? Benchmarking mobile chips remains a nascent science and, much like the PC space, certain benchmarks favour one architecture over another. Intel can readily construct a table whereby unreleased Silvermont looks far better than A15; ARM can surely do the same by picking applications that favour its leaner architecture. We're still some way off from having high-quality usage-based benchmarks that run phones and tablets through everyday tasks - there's little point in providing synthetic numbers that potentially don't tally with real-world experiences.
In due course, Intel will need to impress upon tablet- and handset-makers that its technology is at least as good as the incumbent's. Intel will also draw upon superior manufacturing capabilities and a large marketing chequebook as a means of convincing OEMs to run with new Atom chips. ARM's an R+D house, not a manufacturer, so its strength lies in the partnerships with companies such as Qualcomm, Nvidia and MediaTek, who take on the ARM IP and build it into an eclectic array of chips that power an almost-infinite number of SoCs. ARM will also point to the IP release of the faster Cortex-A57 processor as a sign for optimism, because we imagine the company feels there's little point in comparing shipping technology against an unreleased platform from Intel.
The next six months will prove to be an interesting time for high-performance mobile chips. Intel's coming on heavy with Silvermont-based Atom, ARM's firing back with Cortex-A15 and, in time, Cortex-A57, while, taking just one example, ARM partners such as Qualcomm are sure to muddy the waters by claiming the best balance of performance vs. power through heavily-customised designs that ship under the Krait banner. And if you thought the CPU side of the mobile business was complicated, just wait till you hear about each company's plans for GPU dominance.
