Loading...

The San Jose Nonprofit Defining Machine Learning Small Enough for a Coin Cell

Smart Utilities

Published by IOT San Jose Research & Editorial Team

The San Jose Nonprofit Defining Machine Learning Small Enough for a Coin Cell

The global organization defining and advancing what may be the most consequential category of machine learning most people have never heard of operates out of an office at 75 Headquarters Drive in San Jose. The Edge AI Foundation, known for most of its history as the tinyML Foundation, was founded in 2019 by a consortium including Arm, Google, Microsoft, and Qualcomm specifically to nurture ultra-low-power machine learning technology capable of running on devices operating in the milliwatt range and below, the power budget of a coin-cell battery, not a data center rack. That an organization coordinating research and industry standards for machine learning small enough to run on a microcontroller chose to headquarter itself in San Jose is a fitting detail, since the entire premise of TinyML runs directly counter to the popular imagination of AI as something requiring ever-larger data centers, even as it has quietly become one of the more consequential technology categories shaping how ordinary IoT devices behave.

What Actually Distinguishes TinyML From Other Edge AI

The broader edge AI category already covered elsewhere on this site typically involves dedicated neural processing units running inference models with meaningful compute budgets, often in the multi-watt range, sufficient for tasks like real-time video analysis or generative AI workloads. TinyML operates at a fundamentally different scale: models small enough to run directly on a general-purpose microcontroller, typically consuming power in the milliwatt range or below, without requiring a dedicated AI accelerator chip at all. This distinction matters practically because it determines what class of device can run machine learning at all. A dedicated edge AI chip adds meaningful cost and power draw to a product's bill of materials. A TinyML model running on a microcontroller a device already needs for basic control functions adds essentially no incremental hardware cost, only the engineering effort of building a sufficiently compact and efficient model.

That efficiency requirement shapes TinyML research in ways that differ substantially from mainstream machine learning research, which has spent the past several years scaling models larger rather than smaller. TinyML practitioners work on quantization, reducing the numerical precision a model uses to represent its parameters, model architecture search specifically optimized for extremely constrained memory footprints, often kilobytes rather than gigabytes, and power-aware training techniques that treat energy consumption as a primary design constraint rather than an afterthought measured only after a model is already built.

The Broader Edge AI Foundation Umbrella

The 2019 founding of what was originally called the tinyML Foundation by Arm, Google, Microsoft, and Qualcomm is itself worth noting, since it represents an unusually broad coalition of companies that normally compete directly with each other in the broader AI hardware and software market agreeing to collaborate on shared standards and research for this specific low-power category. That kind of cross-competitor collaboration tends to happen when an industry recognizes a technology category benefits everyone more from shared infrastructure and common standards than from proprietary fragmentation, a pattern also visible in how the Connectivity Standards Alliance brought competing smart home platforms together around the Matter standard. The organization's rebrand to Edge AI Foundation reflects its own recognition that the boundary between TinyML and the broader edge AI category it originally sat apart from has become increasingly blurry as microcontrollers themselves gain more capable, if still modest, on-chip AI acceleration.

Where TinyML Actually Shows Up in Real IoT Products

TinyML's practical applications cluster around always-on sensing tasks where a device needs to make a simple classification decision locally, continuously, without the power budget or connectivity to send data to the cloud for every decision. Presence detection, distinguishing whether a room is occupied without a full camera-based computer vision pipeline, keyword spotting for voice-activated devices that need to recognize a wake word without streaming continuous audio anywhere, and simple anomaly detection on vibration or current sensor data for basic equipment monitoring are among the most common production use cases. Semiconductor companies showcasing low-power FPGA and microcontroller-based AI solutions at tinyML Foundation events have specifically highlighted presence detection and object counting as practical applications that work well within a battery-powered device's power budget, precisely the always-on, low-complexity classification tasks where TinyML's constraints are a genuine fit rather than a limitation.

This overlaps directly with the broader shift toward battery-free and energy-harvesting IoT hardware covered elsewhere on this site, since a sensor running entirely on harvested ambient energy has an extremely limited power budget that rules out anything beyond TinyML-scale inference. The two trends reinforce each other: energy-harvesting hardware innovation makes battery-free sensors practical, and TinyML makes those same power-constrained sensors capable of some on-device intelligence rather than merely transmitting raw data for someone else to interpret.

Business Value for South Bay IoT Product Teams

For hardware companies building IoT products across the South Bay, TinyML offers a meaningfully different cost and capability tradeoff than dedicated edge AI silicon. A product that only needs simple, well-defined classification tasks, is this device being used, has this piece of equipment's vibration pattern changed, is a specific sound occurring, can often achieve that capability without the added bill-of-materials cost, board space, and power draw of a dedicated neural processing unit, making TinyML a genuinely practical default for cost-sensitive, high-volume IoT product categories where a full edge AI chip would be economically difficult to justify.

Implementation Advice for Adopting TinyML

  1. Confirm the use case genuinely fits TinyML's constraints before committing to it, since tasks requiring complex reasoning, generative capability, or high-resolution image analysis exceed what a microcontroller-scale model can realistically deliver.
  2. Budget engineering time for model compression work specifically, since achieving useful accuracy within a kilobyte-scale memory footprint typically requires meaningfully more optimization effort than training an equivalent model without those constraints.
  3. Evaluate existing microcontroller headroom before adding dedicated AI hardware, since many IoT products already include a microcontroller with enough spare capacity to run a well-optimized TinyML model without any additional silicon.
  4. Engage with the broader TinyML community and tooling ecosystem rather than building compression and deployment tooling from scratch, given how much shared infrastructure organizations like the Edge AI Foundation and its member companies have already developed for exactly this purpose.

Common Challenges in TinyML Development

ChallengeDescription
Severe memory and compute constraintsFitting a useful model into kilobytes of memory requires specialized compression techniques most machine learning engineers have not been trained in
Limited task complexityTinyML is well suited to simple classification tasks but cannot handle the complex reasoning or generative capability larger edge AI models provide
Fragmented tooling landscapeDifferent microcontroller vendors and model compression frameworks are not always fully interoperable, complicating cross-platform development
Energy measurement complexityAccurately measuring and optimizing a model's actual energy consumption in the field requires specialized techniques beyond standard software profiling

These constraints connect directly to the battery-free and energy-harvesting hardware innovation and low-power chip design covered elsewhere on this site, since TinyML's practical value depends entirely on operating within the same extremely constrained power budgets those hardware innovations are specifically designed to support.

What's Next for TinyML in IoT

As the Edge AI Foundation's own rebranding from tinyML Foundation suggests, the organization is positioning itself at the center of a broader edge AI movement rather than a narrow technical niche, continuing to run its annual Summit and Research Symposium events to advance the field alongside major semiconductor partners. For South Bay IoT hardware teams, the practical trajectory to watch is how much further model compression and power-aware training techniques can push what were once considered impossible tasks for microcontroller-scale hardware, a boundary that has already moved substantially over the organization's roughly seven years of existence and shows no clear sign of settling into a fixed capability ceiling anytime soon.

Frequently asked questions

TinyML refers to machine learning models small enough to run directly on a general-purpose microcontroller, typically consuming power in the milliwatt range or below, without requiring a dedicated AI accelerator chip, distinguishing it from broader edge AI that often uses dedicated neural processing units with much higher power budgets.

The Edge AI Foundation, formerly known as the tinyML Foundation, is headquartered at 75 Headquarters Drive in San Jose, California, and was founded in 2019 by a consortium including Arm, Google, Microsoft, and Qualcomm.

TinyML works well for always-on, simple classification tasks like presence detection, keyword spotting for voice-activated devices, and basic anomaly detection on vibration or current sensor data, where a device needs a continuous local decision without the power budget to send data to the cloud.

A sensor running entirely on harvested ambient energy has an extremely limited power budget that rules out anything beyond TinyML-scale inference, making TinyML the only practical way to add on-device intelligence to the growing category of battery-free IoT hardware.

Not necessarily. Many IoT products already include a microcontroller with enough spare capacity to run a well-optimized TinyML model without any additional silicon, making it a lower-cost alternative to adding a dedicated neural processing unit for simple classification tasks.

Fitting a useful model into a kilobyte-scale memory footprint requires specialized compression techniques, including quantization and architecture search optimized for extremely constrained memory, that most machine learning engineers trained on larger-scale models have not typically worked with.

Related articles

Top