Ambiq has introduced heliaAOT 1.0, an ahead-of-time AI compiler for the HELIA AI ecosystem that converts trained models into optimized C code for on-device inference. In Apollo510 benchmark tests, the compiler delivered up to 76% lower inference latency and up to 72% lower energy per inference than LiteRT Micro.
The compiler is intended to reduce the processing time, energy and memory required for on-device inference. Ambiq said heliaAOT has been deployed in mass-produced products and that it is discussing deployments in current products with some of its larger customers.
heliaAOT 1.0 supports FP16 and FP32 models without integer quantization. In matched-FP32 benchmark testing across six model families on Apollo510, it delivered approximately 4× faster FP32 inference than LiteRT Micro FP32. Support for 65 operators plus custom operators accommodates more model architectures, while streaming LSTM, GRU and SVDF support enables continuous voice, biosignal and sensor processing.
Key features include median firmware images 29% to 56% smaller than LiteRT Micro; retargeting supported models across 15 Apollo SoCs with a one-line compiler configuration change; and support for continuous voice, biosignal and sensor processing.
Across Apollo510 benchmark workloads, Ambiq reported latency, energy-per-inference and firmware-image reductions of 76%, 72% and 56%, respectively, for health monitoring (ECG, PPG and sleep); 61%, 51% and 29% for vision; 52%, 45% and 47% for voice and audio; and 36%, 27% and 37% for MLPerf Tiny reference models. The tests were run at 96 MHz against LiteRT Micro under matched per-model conditions. Energy figures represent gross MCU-rail energy per inference, and the results are model-level measurements, not battery-life measurements.
Dr. Adam Page, head of artificial intelligence at Ambiq, said: “Customers want to bring more intelligence to their devices without exceeding their power and memory budgets. heliaAOT 1.0 helps them get more from Apollo silicon, deploy a broader range of models, and reuse their development work across products.”
Ambiq said it delivered 17 heliaAOT updates over the past 12 months, expanding operator support from 35 to 65 and Apollo target support from 8 to 15. heliaAOT 1.0 is available, and the company offers benchmark information and evaluation requests at https://ambiq.com/ai/heliaaot/.
