The verified MLPerf™ Tiny benchmark results¹ now show that although GAP9 is dimensioned to deal with much larger problems, it also excels when running very small neural networks (NNs), proving its ability to address the wide range of embedded NN processing tasks required by real life applications.
GAP9 has the lowest energy consumption of all submissions in all four MLPerf™ Tiny benchmarks. In many cases it achieves latency and energy consumption orders of magnitude lower than other devices.
These results are achieved thanks to GAP9’s uniquely scalable architecture and to GreenWaves’ state-of-the-art neural network toolchain, GAPflow, which transforms networks from ONNX and TensorFlow Lite formats into optimised, readable C source code. GAPflow allows for the coexistence of multiple different networks on GAP9 with support for preemption in their execution to meet real time constraints.
GAP9 also has other leading edge computing capabilities that can be combined with Neural Networks. For example, an NN can be used to continuously update the parameters of GAP9’s unique, sample by sample, highly configurable Smart Filtering Unit (SFU) enabling new NN steered Adaptive Active Noise Cancellation or Adaptive Audio Transparency features.
GAP9 provides the most powerful signal processing and AI computing platform for the next generation of hearable and IoT products.
For more information go to www.greenwaves-technologies.com.
 MLPerf™ v1.0 Tiny Closed. Retrieved from https://mlcommons.org/en/inference-tiny-10/ 9th November 2022, entries 1.0-9001 and 1.0-9002. Result verified by MLCommons Association. The MLPerf™ name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorised use is strictly prohibited. See www.mlcommons.org for more information.