heliaCORE

Optimized AI kernels for Ambiq Silicon

Kernel acceleration for Ambiq AI.

ATfE + MVE ready

heliaCORE is Ambiq's optimized neural-network kernel library for Ambiq silicon. This repository delivers it as ns-cmsis-nn: a CMSIS-NN-based kernel layer with Ambiq-tuned operators, CMSIS-Pack delivery, and HELIA integration paths for Apollo-class Cortex-M accelerators.

Getting started Why heliaCORE API

HELIA COREV7.25

200+accelerated ops

40+Field models

53Op types

4Paths

find_package(ns-cmsis-nn REQUIRED CONFIG)
target_link_libraries(app PRIVATE nsx::cmsis_nn)

CMSIS-Pack Zephyr CMake neuralSPOT-X ATfE GCC Cortex-M0+ Cortex-M4 Cortex-M55 Apollo510

What it is

A kernel layer for HELIA AI workloads

heliaCORE is Ambiq’s optimized neural-network kernel layer for Apollo-class devices. In this repository, it is delivered as ns-cmsis-nn: inherited CMSIS-NN-compatible APIs where they apply, Ambiq-tuned kernels where HELIA needs more coverage, and packaging for runtimes, compilers, and firmware builds.

Ambiq workload coverage

Coverage for real Ambiq model graphs

Arm CMSIS-NN provides the trusted foundation for efficient neural-network kernels on Cortex-M. In Ambiq’s HELIA workflows, internal profiling across field-like models also highlighted important Ambiq-specific coverage needs: real graphs spend measurable time in padding, activations, reductions, and other operators around the largest MAC-heavy layers.

heliaCORE broadens Cortex-M accelerator coverage around the CMSIS-NN-compatible foundation so those end-to-end paths stay optimized on Ambiq silicon.

MVE MVE-first where available

Cortex-M55 paths are a primary optimization target, with vectorized kernels where MVE can improve end-to-end latency.

DSP DSP coverage for Apollo-class MCUs

Cortex-M DSP paths remain important on Apollo targets without MVE, so DSP variants stay part of the release surface.

Flow Glue operators count too

Coverage extends beyond obvious MAC-heavy layers to the graph operators that shape real HELIA deployments.

40+ Ambiq field-like suite 40+ models, 53 operator types, 247 unique operators, and 963 operator instances.

Tiny MLPerf Tiny baseline 5 models, 7 operator types, 34 unique operators, and 80 operator instances.

200+ Expanded coverage 200+ accelerated operators for Ambiq silicon, plus additional variants for existing operators.

Explore Cortex-M accelerators Browse operator & kernel coverage

Pick your integration

Four supported ways in

Prebuilt SDK CMake package Link a per-architecture package with the exported CMake target; configure validates CPU and compiler. Start with CMake CMSIS tooling CMSIS-Pack Install the Ambiq pack and select a source or prebuilt variant for your project. Use the pack RTOS module Zephyr Module Enable the module with Kconfig and optionally point it at a prebuilt archive. Wire Zephyr HELIA stack neuralSPOT-X Consume heliaCORE through the NSX CMake graph for heliaRT-based applications. Use NSX

HELIA family

Powering heliaRT and heliaAOT

heliaCORE provides the optimized kernel layer that powers Ambiq’s edge-AI runtime and compiler flows. heliaRT and heliaAOT both rely on optimized HELIA kernels where applicable; heliaCORE is the shared kernel foundation behind those paths.

Runtime path heliaRT Ambiq's optimized LiteRT runtime for Apollo applications, powered by HELIA kernels from heliaCORE where they apply. Open heliaRT Compiler path heliaAOT Ahead-of-time generation for deterministic, low-footprint firmware, using heliaCORE kernels as the shared foundation. See where it fits

Arm CMSIS ecosystem

heliaCORE is built on and distributed for the Arm CMSIS ecosystem, including CMSIS-NN-compatible APIs and CMSIS-Pack delivery. Ambiq-specific additions are intended to ease integration into the HELIA AI platform for Ambiq silicon. For vendor-neutral Cortex-M kernel work, Arm CMSIS-NN remains the upstream ecosystem reference.