SPOT®-tuned hardware-software synergy for ultra-low power AI on Apollo. Choose runtime tools including heliaRT for drop-in, backward-compatible speedups, heliaAOT for compile-time optimization and the smallest, most power-efficient builds, and more. Both runtimes are co-designed with Apollo silicon for kernel optimization, memory planning, and dataflow, so models run efficiently on real devices, not just in benchmarks.
The leak detection industry is evolving. For years, monitoring systems focused primarily on collecting and transmitting data. Sensors…
The dual listing on NYSE and SGX broadens investor access across Asia while reinforcing Ambiq’s long-term commitment to one of the…
Choose heliaRT™ if your product ships with frequent model refreshes and you want backward-compatible, drop-in speed. Choose heliaAOT™ if you’re shipping a locked-down SKU with tight latency, memory, or energy targets and can accept that updates require a rebuild.
Yes — heliaRT™ uses the same .tflite loading flow as TFLM and is fully backward-compatible, so existing TFLM models and workflows carry over with no changes required.
Yes. heliaAOT™ compiles your model ahead of time into a fixed artifact (C/object/binary), so any change to the model requires a rebuild — best suited to a model that’s already finalized rather than one still under active iteration.
heliaAOT™ emits only the code a specific model needs, giving it the smallest memory footprint of the three options, and its compile-time specialization makes it the fastest on Apollo silicon. [Exact percentage figures depend on the model — confirm representative benchmark numbers with engineering before publishing a specific claim.]
Both are SPOT®-tuned specifically for Apollo silicon and built to work across the Apollo lineup. Note heliaAOT™ currently supports a narrower operator/data-type set (int8/int16, no int32) than heliaRT™ or TFLM, so unusual model architectures should be checked against its supported operator list first.
SPOT®-tuned means heliaRT™ and heliaAOT™ are co-designed specifically with Ambiq’s SPOT sub-threshold power platform and Apollo silicon — their kernel optimizations, memory planning, and dataflow are built around how Apollo chips actually run, not a generic microcontroller target. That’s what lets both runtimes deliver real efficiency gains on real devices, not just in synthetic benchmarks.
