Nanoveu Limited (ASX:NVU), a technology company specialising in advanced semiconductor, drone and materials sciences, has disclosed new benchmark results for its ECS-DoT-22 22nm ultra-low-power edge-AI system-on-chip, developed by wholly owned subsidiary Embedded A.I. Systems Pte Ltd (EMASS), showing energy per inference below every published MLPerf Tiny result across all four workloads when using models purpose-built for the chip's deep-learning accelerator.

Key Points

  • New models on ECS-DoT-22 use 17× to 56× less energy per inference than the leading commercial platforms with published MLPerf Tiny results on each workload, including the MLCommons-verified Syntiant Core 2 and the BrainChip AKD1000.
  • Keyword spotting achieves 0.564 µJ per inference and anomaly detection 0.385 µJ per inference, both below one microjoule, measured at 50 MHz with a 0.66 V core supply.
  • The new models exceed MLPerf Tiny reference model accuracy on all four official test sets and deliver latency speedups of up to 9× on keyword spotting, 4.5× on image classification and 3.0× on anomaly detection on the same silicon at 50 MHz.
  • ECS-DoT-22 evaluation kits are available to order following the DigiKey distribution agreement announced on 10 November 2026, with the EMASS IDE and AI compiler due for release in November 2026.

ECS-DoT-22 Energy Per Inference Results Across Four Workloads

The filing discloses energy per inference figures measured on the ECS-DoT-22 evaluation board at 50 MHz with a 0.66 V core supply. Keyword spotting recorded 0.564 µJ with the new model compared with 3.07 µJ for the 2024 reference model, a 5.4× reduction. Image classification recorded 2.67 µJ against 5.5 µJ, a 2.1× reduction. Anomaly detection recorded 0.385 µJ against 0.8 µJ, also a 2.1× reduction. Visual wake words recorded 3.94 µJ against 3.7 µJ, representing a 6% increase in energy consumption for the new model, which the company noted is more accurate and marginally faster than the reference model.

Accuracy and Latency Against MLPerf Tiny Reference Models

According to the filing, all four new models equal or exceed the accuracy of the MLPerf Tiny reference models on the official test sets. Keyword spotting accuracy reached 92.13% against a reference of 91.66% and an MLPerf Tiny threshold of 90%, with latency of 434 µs compared with 3.9 ms for the reference model. Image classification accuracy reached 88.25% against 86.97%, with latency of 1.392 ms against 6.3 ms. Anomaly detection reached an AUC of 0.861 against 0.8595, with latency of 397 µs against 1.2 ms. Visual wake words accuracy reached 86.30% against 84.80%, with latency of 3.474 ms against 5.2 ms.

Comparison Against Published Peer Results and Replicability

The filing states ECS-DoT-22 new model results sit below every published MLPerf Tiny result on energy per inference for keyword spotting, visual wake words, image classification and anomaly detection, comparing against MLCommons-verified submissions from Asygn, STMicroelectronics, Syntiant, Silicon Labs and Renesas, and unverified published results from Ambiq and BrainChip. The company noted that benchmark results presented are company-measured and have not been independently verified or certified by MLCommons. The filing also states that with the EMASS IDE and AI compiler due for release in November 2026, customers and independent third parties will be able to run the same benchmark models and scripts to reproduce the results.