London Calling: Are ARM’s core days numbered?
As we know, success breeds volume and volume provides cash to keep chip development riding the down escalator that is Moore’s Law. Intel has been successful as has ARM. Other architectures are available, as well as multiple implementers. And then there is the application-specific specialization of processors for such things as graphics rendering and other activities.
The system-on-chip era, where the processor is just one part of a larger application-specific chip, changed the business dynamic. Gradually, more and more companies have started to gain time-to-market and cost advantages over their competition by licensing a pre-developed processor core from MIPS, ARC, ARM, Imagination and some others. But of these, ARM has been the most successful by far.
But what happens when everyone is designing their SoCs with cores licensed from ARM? Where is the differentiation? Where is the time to market advantage?
One could argue that there is scope for differentiation in how the system-chip is put together, and that the cores are like the bricks of a house. For example, Nvidia has built a quad-core Cortex-A15 system-chip for mobile applications called Tegra 4, which integrates a fifth processor for low-power standby operation. Samsung’s Exynos 5 Octa is also a quad-core Cortex-A15 design but one that adopts fully the "big-little" power saving approach with a quad-core Cortex-A7. But is that enough differentiation?
If you are using ARM-defined cores and ARM processor optimization packages to develop your chips. you are essentially on the same time-to-market line as everybody else in the world, unless you are one of the lead partners that helps ARM develop the core in the first place. And if you are helping develop the core in the first place maybe you would do better developing that for yourself alone under an architecture license.
This argument could be one of the reasons we are starting to see a rush of ARM architecture licensees coming forward, although it is interesting to note this is predominantly (exclusively?) with companies based in the western hemisphere.
The best known is probably Apple. But Qualcomm has also developed its own ARM-compatible processing architecture called Krait. Broadcom recently announced it had taken architectural licensees for both the ARMv7 and the 64-bit capable ARMv8 instruction set architectures. As ARM moves up to 64-bit computing, the number of architectural licenses seems to have increased with Applied Micro and Cavium both going down the route that allows differentiation at both the processor core and the system-chip level.
Architecture versus cores
However, it must not be forgotten that it takes considerably more resources – time and people – to develop an ARM-compatible processor core than to successfully embed that core and derivatives in a system-chip that wins market share. But for those few companies that do have the vision, resources and the engineering management to start early enough so that they can hit the market with the right chip at the right time, it is a winning proposition.
It is impossible to know how much of Apple’s and Qualcomm’s success with application processors can be attributed to the fact that they now roll their own processor cores. But successful they are. It could be something as simple as this–only the big companies have the resources to justify an ARM architectural license and big companies usually win in the end.
But even if it is something that simple, does consolidation in the semiconductor industry mean that there will be fewer, bigger companies and their need to differentiate themselves from each other will favor the use of ARM architectural licenses over individual core licenses going forward? Such consolidation may even favor the adoption of newer architectures or re-engineered combinations such as MIPS-PowerVR from Imagination.
The answer to the question probably depends on whether you believe in a consolidating or steady-state chip universe. Those in the latter camp will argue that while consolidation may be going on at the top of the industry there is a continual flow of applications and startups emerging to serve those applications, so that the total headcount remains roughly static. Certainly for a startup company licensing an ARM core to get to market quickly would be one cost-efficient way to go – so long as that startup is not trying to go head-to-head against an entrenched giant.
One other thing that will change the dynamic is that Moore’s Law is running out of steam as it becomes more expensive to progress to the next node. It is arguable that as companies are going to spend longer on each given node and are therefore looking to make more differentiation in design as the benefits of cost and power reduction from migrating nodes are more expensive to achieve and less easy to justify.
More differentiation in design may suggest to more enlightened and better endowed companies that an ARM architectural license gives a greater chance of success.
It would be a simplistic counter argument to say that in modern equipment, gadgets and systems it’s the software that makes the difference. It’s true that at compile time the software doesn’t much care whether the ARM processor it is going to run on is the most elegant of proprietary designs or the most widely used vanilla building block. But it is also true that it is processors designed and optimized for a particular application and set of use cases that have the best chance of turning in best-in-class power and performance measurements.
Finally it must also be recognized that the business dynamics around licensing processor intellectual property are different in different industry sectors. So while we may have seen the emergence of architectural licensees and custom ARM processors to address the mobile market and we may yet see the same for server and networking infrastructure, that is not the case in the industrial or automotive sectors. There, where microcontrollers are often the hardware of choice and flexibility, peripheral cores and a great chunk of software are the differentiators, I don’t see much temptation for ARM licensees to do anything but keep taking the cores.