ARM recently introduced the Cortex-A12 processor, designed to cater for mainstream tablets and smartphones of 2014-2015. The processor is due to be integrated into mobile devices by mid-2014. By all accounts, the A12 is overdue, as presently-available Cortex-A9, first commercially introduced in 2009, arguably doesn't have the necessary muscle to rival newer custom-designed solutions from the likes of Qualcomm.
The need for a better chip
Customers who require a cutting-edge mid-range chip from ARM are in something of a quandary until the A12 surfaces, because Cortex-A15, whilst being considerably more powerful, is thirstier for milliwatts. This is one reason why ARM has been advancing the big.LITTLE design, where in a single device the A15 is paired with an energy-efficient Cortex-A7 chip. The device operating system then dynamically switches between the two depending upon load, with A15 handling CPU-intensive scenarios such as web-page loading and the lower-power A7 tasked with menial duties like MP3 playback. This two-processor configuration is currently seen in select Samsung Galaxy S4 smartphones.
big.LITTLE is a nascent solution to a problem that may go right away with A12, whose design is primed for balancing solid performance against power-draw. ARM claims the new chip can offer up to 40 per cent more performance than A9 on a clock-for-clock basis, evaluated on chips hewn on the same process geometry. It is therefore worth exploring the A12's architecture improvements in more detail, to see how well it's set to meet the challenge of powering next-generation mainstream mobile devices.
HEXUS was invited by ARM for technical briefings at the company's headquarters in Cambridge, England. Lead Cortex-A12 CPU architect Fred Piry delineated the improvements ARM has made in A12 over currently shipping A9.
ARM needs to keep continuity with the large install base of A9, and this means it requires a replacement chip that's similar with respect to thermal requirements. However, ARM also appreciates newer technology has to integrate a feature-set that's more suitable for mainstream applications for the next few years - it's worth remembering that A9 was unveiled nearly six years' ago when smartphone/tablet applications were very much in their infancy. This is why the A12's architectural compatibility is far more like A15 and A7's.
Improving performance over Cortex-A9
Going from the top, modernising the design, A12 has provision for a larger physical address space (40-bits) and includes virtualisation support as standard. The chip continues to use ARM's 32-bit-only v7 architecture but there's little reason why we won't see a 32/64-bit v8 version soon; the market is certainly heading that way in the premium space. Processor topology remains intact, that is, each 'cluster' can use from one to four cores.
ARM historically made it optional to integrate a vector-floating-point and NEON unit in the A9, depending upon customers' demands. The firm now believes it makes sense to spend transistor budget and has integrated said unit into the A12, presumably because present and future workloads are likely to place a heavy burden on the floating-point capability of the chip. The unit has been beefed-up, too, as it now features two-way out-of-order (OoO) execution as opposed to in-order for A9. ARM says the NEON-VFP is up to 50 per cent faster on A12 than on A9, calculated on an equal speed and process basis.
Good chip design mandates that operands (data) are kept close to the processor's engine - frequent runs off to main memory are highly inadvisable - and has thus taken the decision to integrate processor-speed L2 cache on the chip. Again, this is costly from a transistor budget point of view but makes sense when looking at simple, effective methods by which to increase performance. The same rationale is true of the double-sized cache lines, stronger translation lookaside buffers (TLB), and wider interconnects, too.
A12 also uses a better branch-prediction unit than A9 - grabbed off the Cortex-A53, we believe - and it'll help mitigate the penalties associated with mispredictions on the longer pipeline used by A12. What's more, on the memory side, which is integral to boosting performance, A12 now has two dual address generation units (AGUs), up from the one on A9, which has previously been seen as an obvious weakness in the older design.
So if the A12 looks so much like the A15 on first glance, how is it different? A12 can be thought of as an interim step between A7 and A15. Keeping the families apart, the top-end shipping processor has a significantly longer pipeline, can issue three instructions concurrently (A12 manages two), and provides deeper buffers.
Cortex-A12 makes a lot of sense right now
All these improvements over long-standing Cortex-A9 do come at a cost. There's every reason to expect frequency-to-frequency power consumption to grow - extra transistors eat juice, after all. Then there's the question of inflated cost, as the A12 is certainly a bigger chip than A9 - damn those extra transistors! - but ARM believes the trade-offs are worth it, especially for the kind of computationally-heavy functions expected to be performed on smartphones and tablets in two years' time.
With A9 now pushed to the limit and A15 considered somewhat thirsty for upper- mainstream devices, ARM's smaller partners would love to see the A12 in full production now, to combat the customised Krait designs from Qualcomm. A12 appears to have good dynamic power range - the ability to scale from energy-efficient usage to high performance - and is highly customisable. It is expected to match the 2GHz-plus frequencies of the incumbent A9 soon after launch.
The Cortex-A12 is a big deal because derivations of the design are likely to power a large chunk of the mainstream mobile market in the future. It appears to be a solid design, capable of scaling down to the position occupied by high-performance A7s and up to low-end A15s. And if partners so choose, future versions of the chip will carry big.LITTLE compatibility with the A7.
A reasonable chunk of ARM's expected revenue is tied to the fortunes of the Cortex-A12. We'll certainly know more by this time next year whether the gamble to switch to another mainstream CPU core has been worth it. Until then, watch this space.