Loading...

Edge AI Chips Get Smaller, Faster and More Efficient

Manufacturing IoT

Published by IOT San Jose Research & Editorial Team

Edge AI Chips Get Smaller, Faster and More Efficient

Texas Instruments says its TinyEngine NPU can cut inference latency by up to 90 times and energy per inference by more than 120 times, compared with similar microcontrollers that lack an accelerator. The first parts to carry it, announced March 10, 2026, include a microcontroller priced under one dollar in volume. Read that sentence twice. A neural network accelerator is now a line item small enough to sit in a chip that costs less than a cup of coffee.

That headline is typical of 2026, and it deserves a careful reading. Edge AI chips are improving along three axes at once, and they are not the same axis. Smaller means lower cost and less silicon. Faster means shorter response times. More efficient means less energy for each decision. Vendors blur the three in launch materials, and product teams that do the same end up disappointed.

Three Kinds of Better

Smaller is about price and area. The sub-dollar TI part is one example. Ceva's NeuPro-Nano NPU, which won an artificial intelligence award at Embedded World 2026, is licensed as intellectual property so that chip designers can drop it into very small devices. RiseLink's BK7259, shown at the same event, combines dual Arm Cortex-M55 cores with an Ethos-U65 micro NPU for 0.3 TOPS, with a keep-alive current below 60 microamps.

Faster is about latency. NXP says the eIQ Neutron coprocessor in its i.MX RT700 microcontroller classifies images 172 times faster than the CPU core alone. Nordic Semiconductor's NPU-enabled SoC, announced in March 2026, claims a fifteenfold speedup over running the same work on a traditional Arm Cortex core.

More efficient is about energy per inference. NXP quotes up to 119 times better energy efficiency than the CPU alone. STMicroelectronics' STM32N6 pairs a Cortex-M55 running up to 800 MHz with a Neural-ART accelerator rated at 600 GOPS and about 3 TOPS per watt. Mobilint's Regulus delivers 10 TOPS at 3 watts, roughly 3.3 TOPS per watt. Ambiq's Atomiq, unveiled at CES 2026, goes after efficiency by running its NPU at sub- and near-threshold voltages.

How to Read the Claims

Most of the dramatic multipliers are measured against a baseline, and the baseline matters more than the number. The table below translates six 2026 claims into what they actually compare.

ChipClassHeadline claimRead it as
TI MSPM0G5187MCU with TinyEngine NPUUp to 90x lower latency, 120x+ lower energy per inferenceVersus similar MCUs with no accelerator; priced under 1 dollar in volume
NXP i.MX RT700MCU with eIQ Neutron NPU172x faster image classification, up to 119x better efficiencyVersus the CPU core alone, not versus a rival NPU
STM32N6Cortex-M55 plus Neural-ART600 GOPS at about 3 TOPS per wattVendor figure for the accelerator under its own test conditions
Ambiq AtomiqSoC with SPOT-optimized NPUAlways-on audio, vision, and reasoning at very low energyAn efficiency-first architecture; compare on your own workload
RiseLink BK7259Dual Cortex-M55 plus Ethos-U650.3 TOPS, under 60 microamp keep-alive, face recognition in under 200 msTargets battery-first devices such as door locks and cameras
Mobilint RegulusVision NPU10 TOPS at 3 wattsAbout 3.3 TOPS per watt on vision workloads

None of this makes the claims wrong. It means they answer different questions. A 172 times speedup against a CPU tells you an accelerator is worth having, not which accelerator is best for your product. The only comparison that settles a purchase is your own model, on each candidate chip, measured for energy and latency together.

Where the Gains Actually Come From

Four design ideas account for most of the progress, and they show up in different places across the table.

  • A dedicated NPU beside the core. TI's TinyEngine, NXP's Neutron, ST's Neural-ART, and Arm's Ethos family all move the repetitive matrix math off the CPU. Arm's Ethos-U85 goes further and supports transformer models at up to 4 TOPS.
  • Lower voltage. Ambiq's approach trades clock speed for operation near the voltage at which transistors barely switch, which cuts energy sharply for work that can tolerate being slower.
  • Keeping data close to compute. Moving numbers between memory and logic often costs more energy than the arithmetic itself. ST's own materials describe computational memory approaches that promise up to 6 times better throughput and efficiency.
  • Brain-inspired designs. Innatera showed its Pulsar neuromorphic microcontroller at CES 2026, aimed at intelligence at the sensor itself.

The Catch: Memory and Model Size

A faster accelerator does not help if the model will not fit. Memory prices rose sharply in late 2025, and running larger models means paying for more of it. This is why model compression has become a first-class skill for firmware teams, covered in our guide to embedded AI and the deployment gap and the primer on TinyML and the Edge AI Foundation.

It also explains why the toolchain is part of the product. ST's STM32N6 launched with official support in Edge Impulse, the San Jose company Qualcomm later acquired, so a model trained there could reach the Neural-ART accelerator through a documented path. When a team evaluates two chips of similar capability, the one with the better compiler and model zoo usually ships first.

What It Means for South Bay Product Teams

Santa Clara County has an unusual number of companies on the supply side of this trend. Cadence in San Jose licenses Neo NPU intellectual property that the company says improves inference per watt by 5 to 10 times over its previous generation. Kinara in Santa Clara built dedicated edge AI processors before NXP acquired it, a story we told in our article on the Kinara deal. SiTime in Santa Clara attacks the timing circuit's share of a device's power budget, covered in our piece on low-power MEMS timing.

For the local teams buying rather than building, the opportunity is mostly about restraint. Many industrial monitoring jobs, such as spotting a bearing that sounds wrong or a motor drawing odd current, fit inside a sub-watt microcontroller. Choosing that tier instead of a larger module keeps unit cost and power down, and it extends battery life in sensors that were previously too costly to maintain. Our look at battery-free IoT hardware shows how far that logic can go.

A Worked Example: Choosing a Chip for a Vibration Sensor

Imagine a team designing a battery-powered vibration sensor for pumps, expected to run five years on a small cell. This is an illustrative scenario. The sensor wakes every minute, samples for two seconds, and classifies the signal as normal, worn, or failing. A small quantized model of a few tens of kilobytes is enough. A part with a modest NPU and a deep sleep current in the microamp range is the natural fit, because the inference burst lasts milliseconds while sleep dominates the energy budget.

Picking a 10 TOPS vision NPU for this job would add cost and sleep current without improving accuracy. Conversely, the same team building a camera that recognizes people at a loading dock needs the larger tier, because image models are far heavier. The lesson is that the right chip is defined by the job and the duty cycle. A buyer who writes both down before opening a datasheet will usually land in the correct class, and will notice quickly when a vendor's marketing is aimed at a different one.

Implementation Advice

  1. Benchmark on your model. Run the same quantized network on each candidate and record both energy per inference and latency, with the memory configuration you will actually ship.
  2. Check the baseline in every multiplier. A claim against CPU-only execution is a reason to use an NPU, not a ranking of NPUs.
  3. Size memory first. Confirm the model, activations, and buffers fit with headroom before choosing the accelerator.
  4. Prefer documented toolchains. Look for official compiler support for your framework and a clear path for updating models in the field.
  5. Plan for the sleep state. For battery devices, measure total energy over a realistic duty cycle, not only the active inference burst.

Common Challenges With Small Edge AI Chips

ChallengeWhy it matters
Marketing baselines differSpeedups are measured against different references, so figures from two vendors are rarely comparable
Model compression effortReaching acceptable accuracy within a small memory budget takes engineering time that schedules often miss
Operator coverageSome NPUs support only a subset of neural network operations, forcing slower fallback to the CPU
Long qualification cyclesIndustrial customers face lengthy validation, which interacts badly with fast-moving chip roadmaps

What Comes Next

Expect small NPUs to learn new tricks. Transformer support on microcontroller-class parts is arriving, and neuromorphic designs are moving from demonstrations toward products. Expect pricing to keep falling too, since TI has said it will integrate TinyEngine across its entire microcontroller range. For the teams using these chips, the practical effect is that on-device intelligence will stop being a feature that needs justification and become the default setting in new sensors. For the wider chip context, see how the AI chip race is reaching IoT devices.

Frequently asked questions

On March 10, 2026, TI introduced two microcontroller families with its TinyEngine NPU, which it says delivers up to 90 times lower latency and more than 120 times lower energy per inference than similar microcontrollers without an accelerator. One part, the MSPM0G5187, is priced under one dollar in volume.

Most large multipliers compare an NPU against the CPU core alone, not against a competing accelerator. NXP's 172 times image classification speedup for the i.MX RT700, for example, is measured against the CPU without the eIQ Neutron coprocessor.

STMicroelectronics rates the Neural-ART accelerator in the STM32N6 at 600 GOPS and about 3 TOPS per watt, paired with an Arm Cortex-M55 core running at up to 800 MHz.

Four ideas stand out: a dedicated NPU next to the CPU, sub- or near-threshold voltage operation as used by Ambiq, keeping data close to compute through in-memory approaches, and neuromorphic designs like Innatera's Pulsar.

A fast accelerator is of little use if the model does not fit in memory. Memory prices rose sharply in late 2025, so teams must size the model, activations, and buffers early and invest in compression and quantization.

Run the same quantized model on each candidate with the memory configuration you plan to ship, and record energy per inference and latency together. Also compare toolchain support, since a mature compiler and documented update path often decide how quickly a product ships.

Related articles

Top