Whether by accident or design, the microcontroller industry made a good collective decision when it standardized on the Arm® M-profile architecture. Today, all the world’s big microcontroller manufacturers use Arm CPU cores, which means that the M-profile architecture is here to stay. Arm can continue to invest in its development, because it can be confident of a steady stream of future license and royalty revenues. From the manufacturers’ point of view, they eliminate the risk of being left stranded if they were to adopt a proprietary architecture that failed to attract widespread support.
In effect, the existence of the Arm M-profile architecture gives MCU makers a way to escape the 1980s consumer’s VHS vs Betamax dilemma.
Might we see a similar sort of convergence to a standard in neural processing units (NPUs) for MCUs?
It is perhaps too early to say which way the MCU market will go. But for embedded device OEMs, the benefits of using an MCU with a broadly adopted NPU architecture apply in the same way as they do to a standard CPU architecture. OEMs using an MCU with a standard NPU get to:
- Avoid the risk of being stranded on a poorly supported architecture that ends up with a shrinking user base. Without broad manufacturer support, NPU architectures will lack the financial firepower required to sustain continual development and updating.
- Draw on a larger pool of AI developers who are already familiar with the chosen hardware platform
- Operate in an environment which is supported by a broad ecosystem of software developers, development tool providers, and machine learning framework providers
- Retain the flexibility to migrate to a difference MCU vendor and carry over their existing AI/ML software to the new hardware
When choosing an MCU for an edge AI device, you’ll definitely need an NPU which can provide good inferencing capability and low-power operation, so that you can run proper AI algorithms on embedded devices which are often powered by a small battery.
But you should also be thinking about the long-term support for the NPU that your MCU contains. It’s too early to identify an industry-standard NPU, in the way that the Arm M-profile is the standard for the CPU in MCUs. It is, though, possible to figure out which NPU architectures are likely to be widely adopted by a cross-section of MCU vendors, and which – often proprietary – NPUs are at risk of being stranded.
In Alif’s case, the existing Ensemble and Balletto AI MCUs include Arm Ethos™-U55 NPUs, and the next generation of products will feature the Arm Ethos-U85. Like the Arm M-profile architecture, the Arm Ethos technology is a third-party product, backed by the resources of Arm and not under the control of any one MCU vendor. The Ethos architecture is well supported by independent software and tools vendors, and by providers of machine learning frameworks such as Executorch.
By choosing an Arm Ethos NPU, Alif is giving customers the assurance that their AI/ML hardware will be broadly supported and enjoy mainstream support from hardware manufacturers and software developers.
That broad support for the NPU is something that every embedded device OEM should look for when making their choice of MCU for edge AI applications.