Key Highlights

  • The ECS-DoT-22 models use 0.564 µJ per inference for keyword spotting, compared with 3.07 µJ for the 2024 reference model.
  • Visual wake words use 3.94 µJ, about 6% more energy than the reference model, although accuracy and speed are higher.
  • The benchmarks are company-measured and have not been verified by MLCommons.
  • A DigiKey distribution agreement allows evaluation kits to be ordered, with the EMASS development tools due for release this month.

Nanoveu Limited (ASX:NVU) has released new benchmark results for its ECS-DoT-22 edge artificial intelligence chip, developed by its Subsidiary Embedded A.I. Systems, according to an announcement dated 12 November 2026. The results cover the four MLPerf Tiny workloads used across the industry: keyword spotting, visual wake words, image classification and anomaly detection. The company says its new models use less energy per inference than leading peers on three of the four workloads, but the results were measured in-house rather than through MLCommons verification.

Energy per inference and processing speed

For keyword spotting, the new model uses 0.564 µJ per inference compared with 3.07 µJ for the reference model published in 2024, a reduction of about 5.4 times. Image classification uses 2.67 µJ against 5.5 µJ, or roughly 2.1 times less, and anomaly detection uses 0.385 µJ against 0.8 µJ. Visual wake words is the exception, with the new model using 3.94 µJ against 3.7 µJ for the reference, about 6 percent more energy.

The company also reports latency improvements on the same chip running at 50 MHz. Keyword spotting is nine times faster, image classification 4.5 times faster, anomaly detection three times faster and visual wake words about 1.5 times faster. Accuracy on each official test set is reported as equal to or above the reference model, and it clears the MLPerf Tiny quality thresholds on all four workloads.

How the comparisons were built

The comparisons extend beyond Nanoveu's own reference models. The company has set its results against published results from Asygn, Syntiant, BrainChip and other vendors, drawing on a mix of MLCommons-verified submissions and unverified results published by Ambiq and BrainChip. Against the lowest published result on each workload, the company reports energy advantages of between 3.3 and 14 times in its energy-delay product comparisons, with larger figures in other comparisons.

The company acknowledges that the methodologies differ. Its reference-model results follow a closed-division approach, while the accelerator-native models are open-division-style results optimised for its own chip. The announcement includes a cautionary statement noting that the comparisons draw on different benchmark rounds and do not establish superiority over every semiconductor available. Energy is measured from the current drawn by the whole chip's core Supply rail, excluding other board components.

Commercial pathway and open questions

The company says ECS-DoT-22 evaluation kits can now be ordered through DigiKey, following a distribution agreement announced on 10 November 2026. The EMASS integrated development environment and AI compiler are due for release in November 2026, and the company intends to make the benchmark models and scripts available so that customers and independent parties can reproduce the results.

Commercial adoption remains subject to customer qualification, integration and procurement decisions, and the announcement does not disclose any design-in wins or volumes. The results are also limited by their in-house measurement, and the visual wake words result shows that the accelerator-native approach does not improve energy use across every workload. Independent replication will therefore be an important test of the claims.

Final Word

Nanoveu's latest benchmark results show lower energy use on three of four workloads and higher speed across all four, measured on its own silicon. The comparisons are detailed, yet they depend on company-measured data and on comparisons with peers that were published under different conditions. The next evidence will come from the release of the development tools, independent reproduction of the benchmarks and any design-in agreements that customers confirm.