Edge AI has a software problem. While the industry has invested heavily in increasingly capable AI hardware, developers still spend significant time optimizing neural network kernels to fully utilize that hardware. heliaCORE™ closes this gap by providing a highly optimized AI kernel library that transforms Ambiq's ultra-low power Apollo SoCs into production-ready Edge AI.
Built as a superset of the open Arm® CMSIS-NN standard, heliaCORE accelerates more than just computationally intensive convolution and matrix multiplication operations. Its key innovation is comprehensive graph optimization across over 200 kernel variants spanning 11 operator categories, including activations, reductions, quantization, reshaping, padding, and data movement. By optimizing the entire neural network pipeline instead of only the headline operators, heliaCORE eliminates performance bottlenecks that commonly limit real-world AI applications.
The results are substantial. Benchmarked on the Ambiq Apollo510 (Arm® Cortex-M55 with Helium™ MVE), heliaCORE matches or outperforms upstream LiteRT for Micro across all 36 benchmarked operators, achieving zero performance regressions and an average 9.4× speedup. Individual operator categories deliver gains of up to 29.8× for reductions, 16.9× for activations, and 11.6× for data movement. Faster execution keeps the processor active for less time, directly reducing energy consumption while extending battery life—critical for always-on, battery-powered Edge AI devices.
heliaCORE also serves as the common optimization foundation for both heliaRT™, Ambiq's LiteRT-compatible runtime, and heliaAOT™, its ahead-of-time compiler, enabling multiple deployment paths from a single optimized kernel architecture. By combining open standards with comprehensive kernel optimization and measurable performance gains, heliaCORE makes high-performance, ultra-low-power Edge AI practical, scalable, and easier to deploy.