ASYGN says ColibryNPU sets MLPerf Tiny energy record

ASYGN announced July 9 that its ColibryNPU microcontroller posted a record-low 22.2 µJ per inference on the MLPerf Tiny v1.4 Visual Wake Words benchmark. The result positions the Grenoble-based company’s RISC-V embedded AI chip for battery-powered vision applications in IoT, wearables, medical devices and other low-power devices.

Why it matters: - ASYGN says ColibryNPU delivers embedded AI processing for vision tasks at less than 1 mW. - The benchmark result targets devices that need long battery life, including IoT products, wearables, medical devices and toys. - A system running one inference per second could operate for more than 3 years on a CR2032 coin cell battery, according to the company.

What happened: - ASYGN announced results for its ColibryNPU microcontroller on July 9, 2026. - The MLPerf Tiny v1.4 results were published on July 7, 2026. - ColibryNPU posted 22.2 µJ per inference on the Visual Wake Words "Human Detection in an Image" test. - ASYGN describes ColibryNPU as a 32-bit RISC-V microcontroller with a neural accelerator.

The details: - ColibryNPU is built for TinyML and ultra-low-power applications. - The chip processes sensor data including video, audio and environmental signals. - The architecture uses near-memory computing, with tightly coupled memory and compute blocks. - The microcontroller can support real-time video processing at several frames per second under 1 milliwatt. - The design uses 8 compute blocks. - Each block performs up to 2 MAC operations per clock cycle. - Peak performance reaches 9.6 GOps/s. - The chip includes a dedicated video processing interface with an Image Signal Processor accelerator. - ASYGN says the system can run on ultra-compact devices powered by small solar panels or standard batteries. - The MLPerf Tiny benchmark is a standardized way to measure latency and energy per inference for embedded AI hardware. - Full benchmark results are available on the official MLCommons benchmark page. - ColibryNPU is available now in an evaluation version. - More information is available through ASYGN contact page.

Between the lines: - The MLPerf Tiny result gives ASYGN a standardized comparison point in a market where performance claims are often hard to verify. - The combination of low energy use and coin-cell battery operation points to devices that can run for long periods without frequent charging or replacement. - ASYGN is positioning ColibryNPU as a platform for embedded vision rather than general-purpose edge AI.

What's next: - ASYGN is now offering ColibryNPU as an evaluation product. - The company is likely aiming to convert benchmark performance into design wins across low-power embedded device categories. - Broader market adoption will depend on how the evaluation version performs in real customer systems.

The bottom line: - ColibryNPU’s record-low MLPerf Tiny energy score gives ASYGN a strong claim in ultra-low-power embedded AI.