Microchip Technology Inc. Announces MPLAB® Machine Learning Development Suite – Setting the Tone for Vendors Looking to Offer Valuable Solution Development Software and Services

Subscribe To Download This Insight

By Tancred Taylor | 4Q 2023 | IN-7081

Microchip continues its initiatives with TinyML, looking to achieve a strong position in a hotly contested market.

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.

 

Microchip MPLAB Release

NEWS


In September 2023, Microchip Technology Inc. announced the launch of its MPLAB® Machine Learning Development Suite. This software suite incorporates into a single ecosystem Microchip’s existing MCU software development portfolio – comprising MPLAB® X IDE (Integrated Development Environment), compilers, and data visualizers to analyze performance of deployed devices and transmit data to multiple Machine Learning (ML) partners’ platforms.

With this launch, Microchip are expanding their capability to offer TinyML development solutions to their customers. This is not Microchip’s first foray into TinyML: previous partnerships and in-house development, for both TinyML software tools and hardware platforms, are regularly announced. The focus of the most recent announcement puts emphasis on making it easier and more efficient to develop and deploy TinyML models. The announcement talks about “streamlined ML model development”, and their ability to automate TinyML development processes (AutoML) by eliminating “repetitive, tedious, and time-consuming model-building tasks” – such that TinyML solutions “can be easily implemented by those with little to no ML programming knowledge”. Many of these are standard fare for new releases for TinyML, but the announcement is nonetheless interesting in focusing not on any new product but on a process. This highlights the growing emphasis which Microchip, like many MCU vendor peers, are putting on highly integrated hardware and software development rather than offering a series of disparate software tools which are of secondary importance to their crown-jewel hardware portfolios.

A Hotbed of Activity

IMPACT


Microchip is not the only company taking an active interest in developing tools to automate ML model development (AutoML) and facilitate TinyML model deployment. Over the past 2 years, there has been a string of efforts, through partnerships and acquisitions, by all-purpose MCU vendors to link up with TinyML and AutoML specialists. Acquisitions include STMicroelectronics’ acquisition of Cartesiam (2021), Renesas’ acquisition of RealityAI (2022), TDK’s acquisition of Qeexo (2023), and Infineon’s acquisition of Imagibob (2023). Acquired TinyML/AutoML vendors tend to be platform-focused vendors, whose products are designed to be hardware agnostic and cover all aspects of the TinyML development and deployment process. The software side is where Microchip too is focusing for the time being with its MPLAB Suite, along with most other all-purpose MCU vendors building TinyML suites in-house – such as NXP’s eIQ, TDK’s Sensor Inference Network (NIF), or Texas Instruments’ EdgeAI Studio.

While many vendors are developing tools in-house, tie-ups with acquisitions and partnerships are key to offering options to customers. TinyML/AutoML platform vendors such as Edge Impulse or SensiML are notable for a range of partnerships with many of the biggest MCU vendors across the IoT ecosystem, offering complementary tools or an alternative for customers to build and deploy their models. For example, Microchip is partnered with both Edge Impulse and SensiML, and additionally counted Cartesiam as one of their partners before its acquisition by STMicroelectronics.

This highlights one of the key aims of vendors offering TinyML solutions: since TinyML will become important to their customers, it becomes an important part of the value chain for which they should develop expertise – but limiting themselves to tools developed in-house can equate to shooting themselves in the foot, for three principal reasons. First, these vendors have primarily distinguished themselves in the past as hardware specialists; while they are developing their software knowledge, other vendors are more specialist. Second, targeted applications vary: TinyML and AutoML require application-specific knowledge to target image analysis, gesture recognition, or anomaly detection, or to use different types of sensor parameters. Third, MCU vendors still see their position in IoT as ‘toolbox’ providers: they provide hardware components and software and services that complement this hardware to allow solution providers or OEMs to address any pain points they might have. This position would be compromised by offering a limited set of tools, and vendors are distinguishing themselves by building broad partner bases with TinyML/AutoML vendors to create as broad an ecosystem as possible. 

The heavy investment of the past two years in TinyML/AutoML software platforms also highlights another interesting dynamic, namely the relative lack of investment by these companies hitherto in the hardware side – such as Field Programmable Gated Arrays (FPGA) or Application-Specific Integrated Chips (ASIC). The reason for this is primarily that all-purpose MCUs are the incumbent solution for the majority of IoT applications and have the cost point and predictable supplies that comes with technology maturity. With the rapidly growing interest and investments around AI chips more broadly, this market for dedicated TinyML hardware is likely one that will see increased scrutiny by incumbent MCU vendors.

ABI Research forecasts that by 2030, the majority of TinyML shipments (75%) will be powered by dedicated TinyML hardware rather than all-purpose MCUs. Vendors such as Edge Impulse take this to heart, developing relationships not only with all-purpose MCU vendors like Microchip and Renesas but also with dedicated TinyML hardware vendors – such as Brainchip, Polyn, Syntiant, and MemryX. To their credit, Microchip in 2022 also announced the launch of their PolarFire® 2 range of FPGAs and SoCs, an updated iteration of its PolarFire® products from 2020. It followed this up in June 2023 with software tools to convert third-party FPGAs to PolarFire® FPGAs– showing the importance of simultaneous software and hardware development to support new TinyML applications. In fact, some of the software tools currently being built by Microchip are explicitly designed to support its growing FPGA range – such as the ability to import external ML models and convert them into lighter versions to be deployed on end nodes.

Microchip, like many of its peers, offers a horizontal toolbox to its customers: it provides the studio environment in which customers can import their models or build new ones, extract relevant features, train models, and compile and deploy them on devices. One growing trend by vendors offering TinyML solutions, however, is to offer pre-packaged solutions targeted at specific use cases, with the aim of offering pre-built data models and high levels of integration between hardware, firmware, and software. One example is Renesas, whose horizontal background with its in-house e2 studio complemented by its acquisition of RealityAI is allowing it to target ‘microverticals’ by offering full services for developing applications from hardware through to software. Renesas currently focuses on two verticals, namely HVAC and motors. Offering full development and deployment services allows MCU vendors to foster a stronger understanding of their customers’ business and evolve their model past hardware distribution with limited service-enablement, and towards becoming full solution providers.

TinyML for the Customer

RECOMMENDATIONS


The market for TinyML and AutoML is young but fast-growing, and the recent interest around all kinds of AI technologies will only accelerate this growth. MCU vendors have been developing their in-house software expertise to target this, and have complemented this with acquisitions and by growing their partnerships with TinyML software platforms. This process is by no means over, and more acquisitions of TinyML software platforms will come as MCU vendors usually experiment in-house first, partner second, and acquire third. Analog Devices’ partnership with Aizip or Silicon Lab’s partnership with Sensory Inc. are potential examples. However, vendors should expect more interest around dedicated TinyML hardware, and this will cause a second wave of partnerships and acquisitions in the coming years.

MCU vendors and TinyML specialists alike should continue to develop their software expertise. This can take several forms, such as increasing the tools for building, training, and compiling models; supporting a greater range of sensor parameters or ML frameworks; building vertically-targeted ‘solutions’; or expanding the analytical insights which their models can provide, to help enterprises move beyond basic anomaly detection into diagnostic and prescriptive actions taken on the end node. One interesting area of development will come from greater sensor flexibility: rather than have individual TinyML models running on devices with different sensor parameters (such as vibration or temperature), customers want more flexibility for integrating sensors into a single package and providing TinyML analytics that can easily adapt to the different signals received without needing extensive reconfiguration. On the hardware front, data logger devices provided by industrial automation vendors, and their more lightweight wireless I/O counterparts offered by the likes of Advantech, Banner Engineering, Ellenex, and others have been growing in popularity and form ideal ground for both sensor sales and TinyML development.

Vendors should also take thought for how to operationalize TinyML technology for their customers. A market dominated by gated or proprietary solutions can cause challenges for adoption such as the need for different ways of compiling and flashing data onto an end node, the ability to access different features, the lack of a single environment for accessing all data models for all equipment and machines, or to visualize data and training models online or offline depending on the supporting vendor’s portfolio. Hardware vendors should consider how to make their hardware configurable and flexible to support features of all different kinds, and software vendors should give thought to how customers can access nodes and systems in the field once the TinyML-enabled devices have been deployed.

Finally, vendors should look to understand their current and potential customer base. MCU and sensor component vendors frequently have big business with machine OEMs, embedding connectivity and sensing within expensive machines which carry out some form of automation process. However, as the IoT world expands, and in particular the focus on condition-based monitoring and Overall Equipment Effectiveness (OEE), a much more diverse range of vendors become potential adopters of the technology. Each of these vendor types have different requirements: while power consumption might be the primary reason for embedding TinyML in a battery-powered sensor node, this might feature less heavily as a requirement for larger mains-powered machinery. Even within industrial maintenance use cases, use cases vary based on the approach to connectivity and sensing, and TinyML solution providers should understand both where their customer base is as well as what each of their requirements are.