Incumbents and Challengers Form a Diverse and Dynamic Ecosystem
|
NEWS
|
We expect demonstrations and announcements around Artificial Intelligence (AI) to be a significant theme at embedded world, especially given recent advancements in ultra-low-power AI processors and novel neural accelerators. There is no shortage of recent announcements from incumbents, as well as startup activity and investments in novel computing architectures, to cater to on-device and far-edge processing with the lowest power envelopes, injecting intelligence into decentralized systems and devices that would previously have relied on cloud or on-premises computing. Recent examples include:
- Infineon’s edge AI platform DEEPCRAFT with support for computer vision models, including model zoo, for a range of Industrial Internet of Things (IIoT) and smart home applications. Progress in automotive use cases has also been seen.
- Microchip’s PolarFire mid-range AI Field Programmable Gate Array (FPGA) offering was expanded with the acquisition of Neuronix AI Labs for the optimization of neural networks for diverse vision-based applications.
- NXP Semiconductor agreed to acquire pure-play edge AI challenger Kinara with a portfolio of AI accelerators and software targeting primarily industrial and automotive sectors.
- Alif Semiconductor integrates Arm’s Neural Processing Unit (NPU) into its Microcontroller Units (MCUs) with the next generation of its Ensemble silicon and expands integration capabilities by partnering with the Edge Impulse Machine Learning Operations (MLOps) platform.
- Syntiant acquired the Consumer MEMS Microphones (CMM) division from Knowles to expand its reach with the prediction that voice will be the interface of the future, targeting diverse consumer devices, particularly in the smart home.
- POLYN, an analog neuromorphic startup, recently introduced an evaluation kit for a tire monitoring solution in partnership with Infineon, the latter of which will supply the System-on-Chip (SoC).
We also see interest gathering around neuromorphic computing, which holds the potential to revolutionize small form factors by mimicking the brain’s neural architecture, enabling ultra-efficient, low-power AI processing in millimeter-sized chips. These systems are popular for form factors like wearable tech, drones, and the Internet of Things (IoT) (ex-Industrial), where size and energy constraints are critical. Key players include POLYN, Aspinity, and BrainChip, targeting primarily consumer applications, and offering intelligent pre-processing, which offloads significant resources from the main processor on power-constrained and battery-powered devices.
Compare and Contrast: Different Strategies Have Emerged
|
IMPACT
|
A pattern starts to emerge when comparing the strategy of incumbents like NXP Semiconductors and Microchip with that of startups like POLYN, Innatera, Aspinity, and others. Incumbents have been fairly slow to move, generally opting for a lower risk approach by targeting a broader set of applications and verticals, as well as, crucially, industrial and automotive markets, which are typically defined by longer lead times and more stringent requirements due to mission criticality. This legacy in industrial and automotive markets, defined by their longer development cycles and their broader remit, may also temper their agility, as smaller startups are able to navigate the relatively more nascent field of AI. Incumbents are also more reliant on acquisitions to bring software capabilities in-house. The acquisitions of MLOps players like Neuronix and Imagimob, by Microchip and Infineon, respectively, are undertaken to internalize powerful optimization tools to support their commercialization of AI/Machine Learning (ML) on power-constrained hardware.
On the other hand, the startups have applied a laser focus to developing highly efficient silicon for a select number of use cases—operating either independently or in conjunction with Microcontroller Units (MCUs)—with software stacks focused around their unique architectures. Depending on their funding, they also have model zoos that customers can leverage to build applications around their own often unique, but always diverse requirements. Furthermore, this group is also targeting primarily consumer applications, like wake word detection for smart home devices and background noise cancelling.
The Significance of 2025, and What We Are Looking Out For
|
RECOMMENDATIONS
|
The challenge for newcomers is convincing Original Equipment Manufacturers (OEMs) to migrate to a new hardware platform. On the other hand, incumbents’ first-generation MCU+NPU offerings are generally viewed as not having been a resounding success. As previously published by ABI Research, these key themes will inform the success of edge AI players across the ecosystem in 2025.
- Winners and losers will emerge, and it is likely that incumbent MCU vendors like Infineon will undertake hardware and software Mergers and Acquisitions (M&A) activity, as seen in the FPGA market with Analog Devices buying Flex Logix, or NXP’s purchase of Kinara.
- We will see the commercialization and mass production of several startups, going beyond the sampling stage, such as Innatera’s T1 neuromorphic microcontroller introduced in 2024.
- There will be more partnerships between incumbents and startups, as seen with POLYN and Infineon on a tire monitoring solution, with incumbents leveraging the performant hardware from newcomers.
- Silicon development cycles to optimize performance for transformer models are well under way, and the next hardware releases will be more targeted toward Generative Artificial Intelligence (Gen AI) at the far edge.
Challenges facing incumbents and new entrants alike include market saturation and tightening customer Time to Market (TTM) expectations. ABI Research forecasts that the following will be signs of future success for market participants:
- Talent shortages limit the broader appeal of edge AI outside of Personal and Work Devices (PWD) like smartphones and laptops. Simplified end-to-end offerings, with comprehensive MLOps toolchains and application, and specific model zoos, are more important now than ever.
- Hardware vendor partnerships will occur with System Integrators (SIs) or solution designers, particularly for those lacking the integration capabilities in-house. This will accelerate the productization of new hardware and spur deployments.
- Verticalized offerings are a sign of maturation and accessibility—this also holds true in the world of cloud AI, where NVIDIA’s enterprise-grade offerings (and software moat) have helped to propel it to its current position of dominance, regarded as a one-stop shop for enterprise AI.
- Voice-based interfaces are set to be a significant driver of growth in the coming years, with transformer-based Small Language Models (SLMs) being integrated into diverse consumer devices. Training platforms to enable end customers with varying levels of data science expertise to launch on hardware platforms are a welcome sign of progress.
- Channels to market and partnerships with vendors whose legacy can be leveraged to inject AI-based intelligence into devices, exemplified by Syntiant’s acquisition of Knowles’ microphone division, are important drivers of growth. These translate to end-to-end solutions with more added value, also offering a differentiator in a highly saturated market.
One major stakeholder not mentioned here is Arm, with its Intellectual Property (IP) forming the backbone of the majority of MCUs today (although less relevant for challengers developing their own silicon). The IP vendor’s in-house NPU offering has progressed, as well as its design services, as it edges closer to becoming a fabless silicon vendor, which has the potential to shake up the dynamics of the market, especially when considering Arm’s reach today. Regardless, work remains to democratize the capabilities among hardware vendors and device OEMs not fortunate enough to have the extensive in-house talent, or funding, to go it alone.