While we still await detailed information on the A8 chip, we know the core parameters: 28nm to 20nm shrink, 1.3GHz to 1.4GHz speed increase, 25% better performance, 50% better GPU performance. This amounts to the smallest update in CPU performance the iPhone has ever seen.
It used to be that a die shrink would enable higher frequencies at the same power level, according to Denard Scaling. So we would have expected a 40% increase in frequency, instead of the 8% we actually got. But this seems to no longer be possible, as also evidenced by the slow progress Intel is making. This means that we are now reaching a significant slowdown in Moore’s Law, so getting better performance will require a different approach. One will need to use the huge transistor budgets now available to create accelerators for bottleneck functions, instead of relying on CPUs to become faster to run the software at a sufficient pace. And it seems that is what Apple has done, with the following uses for the extra transistors:
- 6 instead of 4 GPU cores
- improved CPU, likely including larger caches
- specialized functions for camera and video processing
- Metal provides an interface to the GPU to offload hugely parallel workloads
One wonders what this will mean for the future, and I suspect it will help Apple, as you will need to integrate hardware and software to achieve optimal performance, and it makes Swift an important part of the future, as it allows you to squeeze more performance out of the hardware.
It also means that Intel will have a problem, as they can no longer count on their process prowess to help them overcome the efficiency problems caused by the dated x86 architecture. What they need to do is create a new architecture, optimize the silicon for it, and provide a secondary decoder for x86 that can effectively feed legacy programs to the hardware. This is not easy, but in the current age of billions of transistors, it is feasible to create a dual decoder, where one can power gate the one not in use.