FICUS , home
Reflex Lab
Reflex Lab
Checking…

Reflex Lab · STEP 01 · Results

See the result.
Build Reflex software.

Five measured Reflexes decide in 4.4–54 nJ on Nordic's CPU at an assumed 3.0 V. Run a real example in one click, then evaluate your own data.

See measured results
  • In your browser

    No account. Nothing uploaded.

    • Measured results
    • Example replay
    • CSV profiling
    • Baseline training
  • Open FICUS endpoint

    No account. Graph only, no data rows.

    • Example reduction and its analysis report
  • Approved access

    Sign in. You approve each transfer.

    • Private compiler builds
    • Training and Instinct jobs
    • Board-trace analysis
    • Saved workspaces
Board evidence · Nordic nRF54LM20B DK

See what each Reflex delivers.

Compare the same processing boundary, then inspect quality and resource scope.

V

Calculated, not a voltage-verified measurement. Energy (nJ) = measured charge (nC) × voltage (V). Average power (mW) = energy (nJ) / latency (µs). The capture has no voltage trace; 3.0 V is a scenario, not a verified rail. A GUI maximum is not proof of DUT voltage.

Machine condition

Bearing

MCU + PPK2
Measured mean latency per callLower is better · µs
FICUS1.336 µs
Neuton58.079 µs
TensorFlow Lite100.351 µs
FICUS vs Neuton43.5×faster
FICUS vs TensorFlow Lite75.1×faster

C1 includes each engine’s front and published class, starting with processed float features in RAM. Acquisition and upstream feature extraction are excluded.

Accuracy on the stored validation panel

FICUS100.00%624 / 624 correct
Neuton99.84%623 / 624 correct
TensorFlow Lite88.62%553 / 624 correct

Historical, consumed examples; fresh customer data is needed to establish generalization. These are comparisons of frozen artifacts, not framework-wide rankings.

Deployment memory and input-boundary details

Available size measurements have different scopes

ImplementationRecorded bytesWhat is counted
FICUS16,936 BLinked C1 code and read-only data
Neuton1,259 BModel parameters only; runtime excluded
TensorFlow Lite4,648 BSerialized model only; runtime excluded

These are not equivalent total-deployment footprints. No total-memory advantage is inferred. Per-model peak SRAM and stack are not established.

FICUS persistent state: 0 B. TensorFlow Lite reported arena use: 4,964 B; the benchmark reserves 32 KiB.

What C1 − C0 tells us

FICUS C0: 0.829 µs. FICUS C1: 1.336 µs. Difference: 0.507 µs.

This is a measured boundary difference. It can include caller, preparation, layout and integration effects; it is not automatically a separately measured front that hardware can remove for free.

All five datasets · absolute C1 results

Gross energy and average power are calculated at 3 V assumed supply. Change the supply above to recalculate. Missing C0 measurements remain unavailable.

C1: measured latency and charge; calculated energy and average power; historical panel accuracy.
DatasetEngineLatency
µs / call
Charge
nC / call
Energy
nJ / call
Avg. power
mW
Accuracy
historical
BearingFICUS1.3363.66610.9988.229100.00%
BearingNeuton58.079152.258456.7747.86599.84%
BearingTensorFlow Lite100.351283.378850.1338.47288.62%
ToothbrushFICUS0.5391.4714.4128.184100.00%
ToothbrushNeuton79.839213.661640.9848.028100.00%
ToothbrushTensorFlow Lite108.789307.815923.4458.48824.00%
Magic WandFICUS1.5724.28412.8528.17799.84%
Magic WandNeuton104.770281.241843.7228.05399.19%
Magic WandTensorFlow Lite75.274217.679653.0368.67581.95%
GearboxFICUS6.64818.14854.4448.19098.88%
GearboxNeuton17.78847.200141.6017.96198.48%
GearboxTensorFlow Lite68.045196.685590.0568.67295.75%
TransportFICUS3.3269.60628.8178.664100.00%
TransportNeuton75.431195.258585.7737.76695.09%
TransportTensorFlow Lite152.135424.8991,274.6978.37979.75%

Accuracy uses consumed panels, with different coverage from the timing workloads. TensorFlow Lite Toothbrush is 24%; Transport has 156 recognized-session rows and seven observer rows. These are frozen-artifact comparisons, not rankings at matched accuracy. Deployment sizes and their accounting scopes are available in each Reflex’s details and the CSV; unknown peak SRAM is left empty.

Measurement method and evidence identity

Frozen individual-Reflex firmware on Nordic nRF54LM20, Cortex-M33 at 128 MHz. Two reversed-order rounds; four tasks use 16 equal-weight timing anchors, while Transport uses the stored 163-row trajectory. Accuracy uses the full stored validation panels. Latency comes from MCU timers; PPK2 current integration yields gross charge per call, including idle current during the interval.

Energy ratios assume a common constant supply. Absolute energy requires the DUT voltage. Neither host timings nor RTL estimates are included. These showcase results do not certify a newly uploaded model.

Executive evidence SHA-256: a4a34be2af69e5dea9591a4fc9c5268bf623e9dcb0598485e87ee9e440e599e7

Frozen artifacts compared on one board; historical accuracy panels. The full protocol, caveats and the test bench are on Evidence. How Reflexes compose into an Instinct: Technology.