Skip to content

SPEED vs SIZE Build Variants

Every heliaRT release publishes two optimization profiles. Pick the one that matches your workload's constraints.

The Two Variants

Variant Compiler flags Optimizes for
SPEED (release) -O2 / -Ofast Minimum inference latency
SIZE (release_with_logs / custom -Os) -Os / -Oz Smallest Flash footprint

Debug builds

A third build type — debug (-O0 -g) — is also published for development. Never ship it in production.

Trade-Offs

quadrantChart
    title SPEED vs SIZE
    x-axis "Smaller code" --> "Faster inference"
    y-axis "Lower power" --> "Higher throughput"
    SIZE: [0.25, 0.35]
    SPEED: [0.75, 0.75]
Metric SPEED SIZE
Inference latency ▼ Lower ▲ Higher
Flash usage ▲ Larger ▼ Smaller
RAM usage ~Same ~Same
Best for Real-time audio, always-on wake-word Battery-first, Flash-constrained

How to Select

Download the variant you need from the release bundle:

helia-rt-<tag>/cortex-m55/atfe/release/          ← SPEED
helia-rt-<tag>/cortex-m55/atfe/release_with_logs/ ← SIZE + logging

# SPEED
make ... BUILD_TYPE=release microlite

# SIZE
make ... BUILD_TYPE=release_with_logs microlite

Standard Zephyr optimization flags apply:

# prj.conf
CONFIG_SPEED_OPTIMIZATIONS=y    # SPEED
# or
CONFIG_SIZE_OPTIMIZATIONS=y     # SIZE

Kernel-Level Knobs

The HELIA backend exposes per-kernel optimization overrides:

# Override individual kernels (values: SPEED or SIZE)
CONV_OPT=SPEED
FC_OPT=SIZE

This lets you optimize latency-critical operators (Conv2D) for SPEED while keeping less critical ones optimized for SIZE.

Next Steps