AI Runtimes

SPOT-Tuned AI Runtime

SPOT-tuned hardware–software synergy for ultra-low power AI on Apollo. Choose heliaRT™ for drop-in, backward-compatible speedups; choose heliaAOT™ for compile-time
optimization and the smallest, most power-efficient builds.

heliaRT

heliaAOT

AI Runtimes Highlights

SPOT-Tuned Synergy

Co-designed with Apollo silicon. Kernel optimizations, memory planning, and dataflow tuned to SPOT so models run efficiently on real devices—not just in benchmarks.

Drop-In or Dialed-In

Choose heliaRT for TFLM-compatible, drop-in speedups—or heliaAOT for compile-time control with dials for size, speed, and power.

Smaller Footprint, Faster Starts

heliaAOT emits only what your model needs; heliaRT trims overhead vs baseline TFLM. Result: lean binaries, quick init, more room for features.

Performance per Microamp

SPOT-tuned runtimes squeeze latency and memory while minimizing energy draw, so real-time AI runs comfortably on battery—perfect for wearables, wellness, and smart audio.

AI Runtime Comparisons

Aspect

TFLM

heliaRT(Optimized TFLM)

heliaAOT(Ahead-of-Time)

Primary Fit

Portable baseline
micro-inference

Drop-in, Apollo-tuned
upgrade

Max efficiency & control for production

Performance on
Apollo

Good baseline

Faster via SPOT-tuned
kernels & planning

Fastest via compile-time
specialization/fusion

Memory Footprint

Moderate interpreter +ops

Leaner than TFLM

Smallest—emits only
what the model needs

Deployment &
Updates

Load .tflite at runtime;
very easy swaps

Same .tflite flow;
backward-compatible
with TFLM

Compiled artifact
(C/obj/bin); update
requires rebuild

Maturity / “Battle-
Tested”

Highest (widely used)

High (production-
hardened on Apollo)

Newer (rapidly maturing)

Op & dtype
coverage

Broad ops;
int8/int16/int32

Broad ops;
int8/int16/int32 (TFLM-compatible)

Focused ops;
int8/int16 (no int32)

Optimization
Control

Limited runtime knobs

Apollo-aware planners
& tuned kernels

Most control:
size/speed/power,
layout/schedule

Best For

Rapid prototyping & portability

Products with frequent
model refreshes

Locked-down SKUs
with tight latency,
memory, energy targets