On Tuesday, 30 October, ARM announced its first 64-bit processor offerings, the high-end Cortex-A57 (which will replace the A15) and the Cortex-A53 (which supersedes the A7.)
Given that we have yet to see the first Cortex-A15 and Cortex-A7 devices on the market, it is highly probable that we won’t see hardware based on the A57 and the A53 until 2014, at the very least.
During a presentation by ARM's European VP of Strategy, Ian Drew, a slide showing the relative peak performances of the contemporary ARM cores was displayed, one which gives us an insight into what can be expected from future ARM architectures.
What we already know
According to ARM's slide (above), jumping from a Cortex A8 (65nm) to a Cortex A8 (45nm) will improve relative performance by about 70-75 per cent, thanks to massive increase in clock speed from 600MHz to 1GHz.
Jumping to a dual Cortex-A9 configuration increased the relative performance by around 2.5x compared to the 45nm Cortex-A8. Taking a closer look, we can trace this evolution to three factors: the doubling of cores, a 25 per cent increase in DMIPS performance, and a lower node potentially resulting in the higher clock speed. Yet moving to a 32nm quad-core Cortex-A9 set-up (like the Exynos 4 Quad found on the Galaxy S3) boosts performance by some 25 per cent, a surprising low improvement that we fail to explain.
However, swapping the A9s for two A15s and two A7s nearly doubles the performance at equivalent core count. The A7 is 50 per cent more powerful than the A8, while the A15 is 40 per cent more powerful than the A9. A smaller manufacturing process means that an increase in clock speed is also expected. The A15 is expected to run at up to 2.5GHz while the A7 will run at clock speeds higher than 1GHz on 28nm.
The second generation Cortex-A57 and Cortex-A53 will add another 50 per cent performance increase on a 20nm node, with 16nm and 14nm geometries thought to be in the pipeline. Casting an eye forward, that would almost certainly entail another bump in processing speeds.
So, compared to a 600MHz Cortex A8 (similar to the Exynos 3110 found in the Apple iPhone 3GS, launched in June 2009) etched on a 65nm processor, a 20nm quad-core A57/A53 combo will perform 16.5x faster, while consuming around 30 per cent more energy. But like we said earlier, the first products featuring this turbo-charged architecture aren't expected until 2014, roughly five years after the first Cortex-A8 products appeared.
ARM didn’t post any disclaimer in the presentation deck and we’re almost convinced that the figures were not obtained at constant speeds, so it should be noted that our performance estimates are purely theoretical.
Performance gains on the A57 and A53, the company disclosed yesterday, were down to improvements in the SIMD engine and architecture with a wider pipeline, as well as improvements in the way data is handled in cache.
In other words, expect a bigger performance delta between existing 32-bit products and 64-bit offerings once 64-bit codebase has successfully evolved into the mainstream in a few years' time.