Features
heliaRT keeps the familiar TensorFlow Lite for Microcontrollers programming model while adding Ambiq-focused runtime and kernel optimizations.
Familiar TFLM Workflow
If you already know TFLM, the core concepts stay the same:
.tfliteflatbuffer model inputsMicroInterpreter-based executionMicroMutableOpResolverfor selecting operators- statically allocated tensor arenas
- embedded-focused logging, profiling, and memory tradeoffs
That means existing TFLM knowledge transfers directly to heliaRT, while Ambiq-specific integrations give you a faster path to efficient deployment on Apollo devices.
What heliaRT Adds
- public source integration with
Referenceand openCMSIS-NNbackends - optional HELIA acceleration through Ambiq's private backend module
- build and packaging flows aligned with Ambiq silicon targets
- prebuilt release bundles for faster bring-up
- Zephyr module support alongside source-based integration
- profiling and deployment flows that fit Ambiq developer workflows
High-Level Capabilities
Runtime Compatibility
heliaRT is intended for the same class of microcontroller inference workloads as TFLM:
- quantized inference on memory-constrained devices
- operator-resolver-based builds
- static memory planning
- embedded application integration without dynamic runtime dependencies
Ambiq-Focused Optimization
heliaRT is tuned for Ambiq Apollo platforms and related build flows. Depending on target and configuration, this includes:
- optimized int8 and int16 operator paths
- DSP and MVE-aware kernel implementations where available
- release bundles prepared for supported Ambiq-oriented build matrices
Flexible Integration Paths
Users can start in the way that best matches their project stage:
neuralSPOTfor quick profiling and deployment withns_autodeploy- Zephyr raw modules for source-visible integration
- Zephyr prebuilt bundles for fast-start integration
- direct source builds for custom environments
Choosing a Path
Use Zephyr setup if you are integrating heliaRT into a west workspace.
Use neuralSPOT setup if you want the fastest path to evaluate and profile a model on Ambiq hardware.
Use Examples when you want working integration patterns rather than general setup guidance.