New process, new product - ARM details its upcoming Cortex-A72 microarchitecture
by
, 05-02-2015 at 03:54 AM (1059 Views)
Earlier this year, ARM announced its Cortex-A72 — a new custom microarchitecture from the CPU designer that builds on and refines the 64-bit Cortex-A57. Ordinarily it takes up to 24 months for new ARM cores to come to market, after the company announces a new CPU design. But Qualcomm has*told us to expect Cortex-A72 cores by the end of the year. If true, that would make this one of the company’s fastest CPU ramps, ever — so what can the new core do?
If ARM hits its targets, quite a lot.
New process, new product
The Cortex-A72 is based on the Cortex-A57, but ARM has painstakingly refined its original implementation of that chip. The company is claiming that the A72 will draw 50% less power than the Cortex-A15 (a notoriously power-hungry processor) at 28nm and 75% less power at its target 16nmFF+ / 14nm process node. Compared to the Cortex-A57 at 28nm, ARM still expects the A72 to draw 20% less power.
ARM is supposedly aiming for the Cortex-A72 to be capable of sustained operation at its maximum frequency, which is a topic we touched on yesterday when covering the Snapdragon 810’s throttling problem. The CPU is targeting improved performance of 1.16x to 1.5x over the Cortex-A57, clock-for-clock. Making this happen required revamping the branch predictor, cutting misprediction by 50%, and a 25% reduction in speculation power consumption. The chip can also bypass its branch predictor completely in circumstances where it is performing poorly and save additional power in the process.
The Cortex-A72 is still capable of decoding three instructions per clock cycle, but apparently adds some instruction fusion capability to increase efficiency. Each of these components has been power-optimized as well. AnandTech reports that ARM’s dispatch stage can break fused ops back into micro-ops for increased execution granularity, effectively turning a three-wide decoder into a five-wide machine in some cases.
ARM is also amping up its game in SIMD execution units. Instruction latencies have been slashed, pipelines shortened, and cache bandwidths boosted. There are no huge changes in organization or capability, but the CPU core should see significant improvements thanks to these adjustments. ARM has even managed to shave off some die size — the Cortex-A72 is supposed to be about 10% smaller than the Cortex-A57, even on the same process.
Ars Technica reports that according to ARM, the Cortex-A72 can even beat the Core M in certain circumstances. Such predictions must be*taken with a grain of salt — they assume, for example, that the Core M will be thermally limited (we’ve seen that this can vary depending on OEM design). Tests like SPECint and SPECfp tend to be quite dependent on compiler optimizations, and while the multi-threaded comparison is fair as far it goes, ARM is still assuming that the Cortex-A72 won’t be thermally limited. Given that all smartphones and tablets throttle at present, the company will need to prove the chip doesn’t throttle before such claims can be taken seriously.
All the same, this new chip should be an impressive leap forward by the end of the year. Whether it’ll compete well against Apple’s A9 or Qualcomm’s next-generation CPU architecture is another question.
More...