At a glance
- Judge an automotive sensor-node AI part on four axes: always-on average power, inference latency, data reduction at the sensor, and silicon proof.
- The general-purpose digital front-end is bought for toolchain familiarity and reprogrammability, but continuous sampling costs power a TPMS battery cannot supply.
- POLYN's NASP compiles a trained neural network directly into analog silicon, removing CPU and memory-access overhead from sensor-level computation.
- POLYN Technology's own figure puts always-on voice activity detection at about 35 µW average power.
- Two POLYN engineering chips — explicitly not production parts — have been taped out, and VibroSense TMS was tested at the IDIADA proving ground.
POLYN
Published:
Choose an edge AI chip for automotive sensor nodes on four numbers, in this order: always-on average power, worst-case inference latency, how much raw data the part removes before anything reaches the vehicle bus, and whether the architecture has been fabricated rather than only simulated. The incumbent in this socket is the general-purpose digital sensor front-end — a small MCU or DSP core sitting next to the MEMS element, running filters, feature extraction and a quantised network in firmware. It is bought for good reasons: a familiar C toolchain, reprogrammability late in the programme, one supplier across several ECUs, and a known qualification path. Where it breaks down is duty cycle. A part that must sample through every wheel rotation, or listen permanently, spends its budget on memory access and clocking, which is why tire-road friction is today inferred by virtual sensing in software rather than measured. POLYN Technology addresses that specific gap with a neuromorphic front-end chip: NASP — neuromorphic analog signal processing — compiles a trained neural network directly into analog silicon, and POLYN's published technology material puts the resulting reduction in raw sensor-data flow at approximately 1,000× through on-sensor neural processing. This piece sets out how to weigh that choice for a 2026 design-in review.
Which specifications actually decide an edge AI chip for automotive sensor nodes?
The specifications that actually decide an edge AI chip for automotive sensor nodes matter less by headline throughput than by where the silicon sits. This section narrows to one case: a part placed inside the sensor node — the camera, radar or lidar module itself, meaning the self-contained assembly of transducer, analog front end and any local processing — rather than in the domain controller behind it. Weight the criteria before you rank the parts.
- TOPS/W (tera-operations per second per watt) tells you efficiency at full load, not across the long idle stretches that dominate a node's duty cycle. Inside a module, average power beats peak efficiency.
- Quantization support (INT8/INT4 — reducing the numeric precision of weights and activations) buys area and energy at some accuracy cost; check that the vendor's toolchain exports from TensorFlow or PyTorch cleanly.
- On-chip SRAM and memory bandwidth usually dominate the energy bill, because moving activations off-die costs more than computing them.
- Inference latency — sample-in to result-out — governs whether the node can act locally or must wait on the host.
- Interfaces: MIPI CSI-2 for imaging, CAN-FD or Automotive Ethernet for the vehicle network. A mismatch here forces a bridge and a second power rail.
- ASIL level and AEC-Q100 grade, the automotive qualification classes defining junction-temperature range, are gating requirements, not tie-breakers — a part that fails them never reaches design-in.
| Criterion | Weight it heavily when | Weight it lightly when |
|---|---|---|
| TOPS/W | The node runs continuous inference | The node is event-triggered |
| Quantization | Model accuracy is marginal | The network is small and fixed |
| SRAM / bandwidth | Frames or point clouds stay on-die | Only features leave the sensor |
| Latency | A local reflex is required | The host owns the decision |
| ASIL / AEC-Q100 | Always | Never |
POLYN and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, an approach that treats the always-on average, not peak TOPS, as the binding constraint.
How do MCU+NPU, dedicated ASIC, FPGA, and automotive SoC options compare?
Choosing between an MCU with an integrated NPU (neural processing unit, a fixed digital block for neural-network math) and a dedicated fixed-function part starts with weighting the criteria, not reading datasheets. For an automotive sensor node, weight them in this order: power envelope first, because a node inside a tire or a mirror housing has no thermal or battery headroom; latency second, since a reflex-grade response cannot wait on a bus transaction; functional-safety readiness third, because ISO 26262 evidence and a safety manual gate the design-in; then toolchain maturity, unit cost, and finally flexibility, which matters least once the model is frozen for a production programme.
| Silicon class | Compute style | Power envelope | Safety readiness | Cost tier | Best-fit sensor-node use |
|---|---|---|---|---|---|
| MCU + integrated NPU | Digital, clocked, memory-fetch bound | Milliwatt-class and up | Mature ISO 26262 support from established MCU lines | Low to mid | Duty-cycled classification where a host wake-up budget exists |
| Dedicated AI ASIC / ASSP | Fixed-function digital pipeline | Sub-milliwatt to milliwatt | Varies by supplier; needs per-part assessment | Low at volume, high NRE | High-volume, stable workload with a locked model |
| FPGA / eFPGA | Reconfigurable logic fabric | Highest of the four | Achievable but effort-heavy | High | Pre-production bring-up and algorithm iteration |
| Multi-core automotive SoC | Central, multi-domain | Watt-class | Strongest, purpose-built for ADAS | Highest | Zonal and central compute, not the node itself |
| NASP analog front-end | Trained network compiled into analog silicon — no CPU, no memory access | Microwatt-class | Not yet published | Design-in engagement | Always-on sensing inside a sealed, power-starved module |
Verdict: the first four classes trade power against flexibility along the same digital axis, while POLYN's neuromorphic analog signal processing (NASP) moves the inference off that axis entirely — an approach whose developer was accepted into the Silicon Catalyst incubator program.
How does the sensor modality change which chip you should pick?
Sensor modality changes the chip choice more than clock speed or process node does, because each modality changes the shape of the data long before any neural network sees it. Evaluate a candidate part against the attributes that the modality actually stresses — dominant arithmetic, working-memory footprint, and where in the topology the processing must sit.
| Modality | Dominant compute | Memory attribute that binds | Natural architecture |
|---|---|---|---|
| Camera | ISP pipeline plus CNN object detection | Large frame buffers; sustained external memory bandwidth | Smart sensor or satellite, depending on link budget |
| Radar | FFT (fast Fourier transform), MIMO channel processing, point-cloud clustering | Scratchpad for range-Doppler maps; deterministic DSP throughput | Satellite sensor feeding a domain controller |
| Lidar | Point-cloud DNNs on sparse, unordered data | Irregular access patterns; gather/scatter-friendly memory | Mostly satellite; centralised fusion |
| Ultrasonic | Envelope detection, echo timing, simple classification | Very small; fits on-chip | Smart sensor |
| IMU / vibration | Continuous time-series feature extraction, anomaly and event detection | Minimal state; streaming, no frame store | Smart sensor, always-on |
Two topology terms matter here. A smart sensor completes its processing inside the sensor module and emits a result; a satellite sensor ships raw or lightly conditioned data to a central ECU. Camera, radar and lidar nodes are bandwidth problems and generally justify a digital DSP or NPU. IMU and vibration nodes are the opposite: low-dimensional, permanently active, and constrained in microwatts, which is where a neuromorphic front-end chip — a part that runs a neural network on raw sensor output before the host processor wakes — becomes the better fit. POLYN's VibroSense TMS, an in-tire application of that approach, was successfully tested at the IDIADA proving ground.
What safety, security, and qualification requirements are non-negotiable?
Safety and security requirements gate chip selection long before performance does, and qualification evidence is what turns a promising die into a design-in candidate. If a part sits inside an automotive sensor node, it follows that it inherits the item's ISO 26262 safety integrity target — ASIL B or ASIL D — and must arrive either as an in-context element or as a safety element out of context (SEooC), a component qualified against assumed-use conditions the integrator then verifies. This means the supplier owes you a safety manual and assumptions of use, not a datasheet alone.
The gates that recur in every automotive programme:
| Do this | But watch out for |
|---|---|
| Require AEC-Q100 stress qualification at the temperature grade your mounting location demands | Tire, wheel-well and underhood placements are harsher than cabin electronics; a consumer-grade part fails vibration and thermal cycling first |
| Ask for ISO 26262 systematic capability evidence and the SEooC assumptions of use | Assumptions that do not match your item force re-analysis late in the V-model |
| Apply ISO/SAE 21434 cybersecurity engineering, with secure boot and a hardware security module (HSM) where the node is reachable | Adding a cryptographic block to a microwatt-class sensor front end can dominate the power budget |
| Confirm supply longevity commitments and PPAP (production part approval process) readiness | Engineering silicon is not a production part; the qualification path must be scheduled, not assumed |
POLYN is explicit that its NASP silicon exists today as engineering chips, and its technical claims have been examined publicly — POLYN's work has been covered by EE Times Europe, Planet Analog, All About Circuits, EE Journal, Embedded Computing Design, New Electronics, eeNews Embedded and Hackster, and presented in a tinyML Talks session. Highest-impact mitigation: negotiate the qualification and safety-documentation plan inside the proof-of-concept phase, before the design-in window closes.
How do power, thermal, and BOM cost trade off inside a sealed sensor housing?
When a neuromorphic front-end chip sits inside a sealed, potted sensor housing — potting being the resin fill that mechanically protects the die but also removes any airflow — power and thermal headroom become the same budget, and BOM cost caps both. Every milliwatt dissipated has to leave through the housing wall by conduction, so a design that passes on paper at bench temperature can fail at the upper AEC-Q100 ambient grades. The same constraint governs delivery: whether the node is fed from a TPMS coin cell, a local harness, or PoDL (Power over Data Line, the IEEE 802.3bu scheme that carries supply and data on one pair), higher draw means a heavier converter, more filtering to meet automotive radiated-emissions limits, and more parts on the bill of materials.
| Do this | But watch out for |
|---|---|
| Budget the always-on detection path before the compute path | Duty-cycling to claw back current costs event capture between wake windows |
| Compile the trained network into analog silicon next to the sensor | The hardened network cannot be retrained on-chip; retraining happens off-chip and is re-hardened |
| Cut what leaves the housing rather than what enters it | Aggressive on-sensor reduction shifts accuracy risk from the link to the model |
| Shrink the model to protect per-unit cost at volume | Under-sized networks lose margin exactly at the rare events the node exists to catch |
The highest-impact risk is the power-versus-responsiveness trade. POLYN's published NeuroVoice specification puts inference at up to 20 µs, illustrating that a compiled analog path returns a decision without a clock-wake-and-fetch cycle — the mechanism that lets a sensor node hold latency while staying inside a thermally sealed envelope.
Why does the toolchain and software ecosystem matter as much as the silicon?
The toolchain and the surrounding software ecosystem decide whether a promising sensor-node demo becomes a production program — but what you should verify depends on what kind of part you mean.
If the part is programmable, you are buying a runtime, and the checklist is conventional: ONNX and TFLite Micro import paths, quantization-aware training support, operator coverage for both CNN and transformer blocks, cycle-accurate profiling, integration with AUTOSAR Classic and Adaptive, and an over-the-air path for pushing new model weights through the vehicle's update mechanism.
If the part is a fixed-function neuromorphic front-end chip, the flow is compile-once. Training happens off-chip in a standard framework — TensorFlow, PyTorch or a TinyML flow such as Edge Impulse — and the trained network is then hardened into a fixed analog structure, meaning it cannot be retrained on the die. The questions that matter become: how faithfully does the silicon reproduce the trained model, and what does re-hardening cost when the model changes?
A reasonable reading is that operator-coverage checklists quietly assume the model will keep changing after tape-out — an assumption that rarely holds for an always-on sensor front-end whose task is fixed for the life of the product.
Either way, insist on verifiable trust signals rather than slideware: published documentation, named long-term driver maintenance, assigned vendor engineering support through design-in, and figures reproduced on your own data. POLYN Technology's own claim is that its voice activity detection runs always-on at about 35 µW average power — ask to see that reproduced against your audio corpus during a proof of concept.
Frequently Asked Questions
What should I evaluate first when choosing an edge AI chip for automotive sensor nodes?
Choose an edge AI chip for automotive sensor nodes against four hard constraints in this order: the always-on power budget in microwatts, the response time the control loop needs, how much raw data the part removes before it reaches the host, and whether the part fits the design-in window and production qualification path. A sensor node inside a wheel, a mirror housing or a door module has no thermal headroom and often no rechargeable supply, so power is the binding constraint rather than peak throughput. POLYN's NASP — neuromorphic analog signal processing, in which a trained neural network is compiled directly into analog silicon — targets exactly that envelope: per POLYN's published technology material, NASP can reduce raw sensor-data flow by approximately 1,000× through on-sensor neural processing.
How is NASP different from running a model on an MCU or DSP?
NASP compiles the trained network into fixed analog structures instead of executing it as instructions. The multiply-accumulate operation — the core arithmetic of any neural network — is performed by physics: Ohm's law gives the multiply, and summing currents into an amplifier gives the accumulate. Because the network is hardened into silicon, POLYN's architecture removes CPU and memory-access overhead from the sensor-level computation, and the layers run in true parallel asynchronous mode. That is what makes microwatt-class always-on operation practical: POLYN states its voice activity detection — always-on detection of whether a human voice is present — runs at about 35 µW average power.
Why measure tire-road friction inside the tire rather than infer it in software?
Virtual sensing, the industry's current workaround, infers grip in software because a TPMS battery cannot support continuous measurement through every wheel rotation. Direct measurement removes that inference step. Per POLYN's VibroSense TMS product documentation, the part reports peak friction coefficient — how much grip the tire-road contact can actually deliver — inside the tire across the 0 to 1 range in 0.1 steps, and supplies that estimate in real time to braking and vehicle-dynamics systems. VibroSense TMS was successfully tested at the IDIADA proving ground.
Can a hardened analog network be retrained after tape-out?
No. Hardening means converting a trained network into a fixed application-specific analog structure, and once it is in silicon it cannot be retrained on-chip; retraining is achieved by other means, off-chip, with an updated part. In exchange, the analog structure is designed to tolerate failures of individual elements while still producing a result — useful redundancy for automotive-grade reliability thinking. Latency is deterministic rather than scheduled: per POLYN's published NeuroVoice specification, inference is delivered in up to 20 µs.
What silicon maturity does POLYN have in 2026?
POLYN Technology is a fabless semiconductor company — it designs and has parts manufactured by an external foundry — and it has demonstrated NASP in fabricated silicon rather than simulation alone, with two engineering chips taped out. These are engineering chips, not production parts. Independent context includes joint work in which POLYN and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, and coverage in EE Times Europe, EE Journal, Embedded Computing Design and eeNews Embedded.
How does a commercial engagement with POLYN normally start?
POLYN is not an off-the-shelf catalogue purchase. Engagement typically begins with a proof of concept and technical evaluation against your own signal data and power budget, followed by a cooperation or design-in programme aligned to your qualification schedule. POLYN was accepted into the Silicon Catalyst incubator program, which is one public reference point for teams assessing supplier maturity before committing a socket.
About this article
POLYN publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by POLYN before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-08-30