The folks at ARM today announced a new suite of technologies aimed at future smartphones coming in the 2016 time frame. We didn't get a ton of details about the underlying architectures involved, but ARM announced successors to its current top-of-the-line CPU and GPU cores, along with a new north bridge to glue them together.
The CPU core is called the Cortex-A72, and it supplants the Cortex-A57 as ARM's fastest CPU core. The Cortex-A72 will support big.LITTLE multi-architecture power-saving configurations by pairing up with the "little" Cortex-A53 core. Both of these cores support the 64-bit ARMv8-A instruction set.
Performance comparisons with past CPU cores are almost always difficult to make in this context, particularly because ARM itself doesn't make chips or consumer devices. The firm offered one general performance estimate, which features the eye-popping number you can see above: a claimed 3.5x performance improvement for the Cortex-A57 over the older Cortex-A15.
Let's contextualize that number. The Cortex-A15 used for a baseline is built into a chip manufactured on a 28-nm process. This chip presumably shipped in a smartphone introduced last year. The next step up the ramp is a Cortex-A57 core built on a 20-nm process—almost exactly like the Exynos 5433 chip used in the Korean version of the Galaxy Note 4, whose performance we've recently measured.
ARM reckons the A57 in this scenario is 1.9 times as fast as the Cortex-A15. Now, step up to the Cortex-A72 microarchitecture and step down to a 16-nm fab process that incorporates FinFETs (or 3D transistors). The combination of process tech and architectural improvements should yield a 3.5x increase over the Cortex-A15. That should work out to a CPU performance improvement of about 84% over the Exynos 5433.
ARM didn't reveal the degree to which process tech improvements, rather than architectural changes, contribute to the expected performance gains. However, we have some clues. The firm is offering pre-baked physical IP products (or POPs) for TSMC's 16-nm process, and it told us the POP enables mobile implementations of the Cortex-A72 to reach clock speeds of up to 2.5GHz. The Cortex-A57 cores in the Exynos 5433 today run at 1.9GHz in the Galaxy Note 4. Pushing from 1.9GHz to 2.5GHz yields a 32% increase in clock frequency. Presumably, the remaining ~50% of performance uplift that ARM expects out of the A72 should be the result of architectural changes that allow higher instruction throughput in each clock cycle.
The Cortex-A57 is a fairly formidable performer in the mobile space, so a ~50% IPC gain would be pretty dramatic. Then again, this slide is about "sustained" performance, so we may be looking at some predominantly power-limited scenarios where IPC doesn't have a ton of bearing on the outcome.
Of course, performance improvements can also be turned into power savings. ARM expects the A72 to use 75% less power than the A15 does when processing the same workload. The power savings are even more pronounced if big.LITTLE is deployed to good effect. Again, these numbers have process tech improvements factored into them.
ARM says it already has more than 10 licensees for the Cortex-A72.
I suspect the performance numbers for the Cortex-A72 above also factor in another important variable: ARM's new CCI-500 north bridge, which offers twice the peak bandwidth of today's CCI-400 and promises a 30% improvement in CPU memory throughput. That's worthy of note since we found that the CCI-400 in the Note 4 is kind of sluggish in directed tests of memory bandwidth compared to solutions from Apple and Qualcomm. ARM also notes that the CCI-500's higher internal bandwidth helps enable 4K displays, which is apparently going to be a thing on future smartphones regardless of what you or I might think about it.
The final piece of ARM's new suite of smartphone tech is an upgraded GPU known as the Mali-T880. ARM claims the T880 will achieve up to 1.8x the graphics performance of last year's mobile devices or a 40% energy reduction for the same workloads. The point of comparison here is the Mali-T760 GPU that we recently tested, but I suspect the assumptions baked into these numbers involve 28-nm silicon, not the 20-nm chip we tested. Of course, GPUs can be wider and narrower, so doing a direct comparison like we have with the A57 and A72 won't really fly without more specific information.
The T880 presumably sits above the Mali-T860 and friends, announced a few months back, at the top of ARM's graphics product stack. All of these GPUs are meant to pair up with the V550 and DP550 video and display controllers, which have native support for H.265 video processing and 4K resolutions.
ARM tells us it will reveal more architectural details about its new offerings later this year, so I expect we'll have some better answers to some of our questions before too long. In the meantime, all of this new technology is available now for ARM's customers to license.
