SAN FRANCISCO, September 17, 2025 — Leads & Copy — MLCommons® has announced the results of its industry-standard MLPerf® Tiny v1.3 benchmark suite, designed to measure the performance of small neural networks in an architecture-neutral, representative, and reproducible way. These networks, typically under 100 kB, process data from sensors like audio and vision for endpoint intelligence in low-power devices.
Version 1.3 includes a new one-dimensional depthwise separable convolutional neural network (1D DS-CNN) test. Trained on sequential data such as sensor readings or audio waveforms, the 1D DS-CNN identifies signals, triggers, or threshold events in real-time data streams. The new test assesses wake-word recognition in a continuous audio stream, exercising capabilities like low-power idle, rapid wake-up, data ingestion, and feature extraction. This capability opens new opportunities, including speech enhancement, real-time translation, and industrial monitoring.
This version also introduces a new, open-source test harness to simplify access and execution of the benchmark suite. The release includes 70 results across five benchmark tests from Kai Jiang, Qualcomm, ST Microelectronics, and Syntiant, including 27 power results. Five hardware platforms were benchmarked for the first time.
MLCommons is an open engineering consortium with over 125 members. It develops benchmarks and metrics for better AI, helping to evaluate and improve AI technologies’ accuracy, safety, speed, and efficiency.
For additional information on MLCommons and details on becoming a member, please email participation@mlcommons.org.
Contact: press@mlcommons.org
Source: MLCommons
