Skip to content
FICUS by Numascale, home
FIG. E1Evidence

Evidence · November 2026

Three studies. Every measurement.

Measured on Nordic’s nRF54LM20, with every method and caveat below.14

Nordic average · seven workloads

Less energy. Less time.

Seven workloads against Neuton and TensorFlow Lite.

ReferenceLess energyFaster
Neuton36.7×40.2×
TensorFlow Lite84.1×86.0×
Average detailsPer-workload ratios
How many times slower, and how many times more energy, each reference needs than FICUS, per workload
WorkloadNeuton timeNeuton energyTF Lite timeTF Lite energy
BearingC143.5×41.5×75.1×77.3×
ToothbrushC1148.1×145.3×201.8×209.3×
Magic WandC166.6×65.6×47.9×50.8×
GearboxC12.7×2.6×10.2×10.8×
TransportC122.7×20.3×45.7×44.2×
Eye trackingenergy-first · X/Y41.2×29.1×161.0×135.9×
MNIST digitslite12 vs Neuton · P3 vs TF Lite157.3×148.3×635.6×555.3×
Geometric mean40.2×36.7×86.0×84.1×

Computed · equal-weight geometric means1

FIG. E2Study 1 · five Reflexes

Study 1 · five Reflexes

4.4–54 nJ per decision.

Reflex software finishes first against Neuton and TensorFlow Lite on the nRF54LM20 CPU.

29.1×

less energy than Neuton

30.4×

faster than Neuton

0.54–6.65µs

per decision

Measurement setupMeasured · nRF54LM20B DK
The Nordic nRF54LM20 DK development board used for all three studies
Board
Nordic nRF54LM20B DK
Processor
Arm Cortex-M33 at 128 MHz, CPU only
Boundary
C1: processed float features in RAM → published class
Instrument
Nordic PPK2, about 1 kS/s
Protocol
64 phases, 784 windows, frozen artifacts
Energy
Measured charge × assumed constant 3.0 V
Datasets
5 tasks, 6,288 rows
Evidence
Nordic evidence v6.3, 4 Nov 2026
FIG. E3Study 1 · per Reflex

Energy per decision · shared linear scale · lower is better

Bearing · Industrial condition monitoring

41.5× vs Neuton

FICUS
11.0 nJ
Neuton
457 nJ
TF Lite
850 nJ

Toothbrush · Wearables and personal care

145.3× vs Neuton

FICUS
4.4 nJ
Neuton
641 nJ
TF Lite
923 nJ

Magic Wand · Wearables and personal care

65.6× vs Neuton

FICUS
12.9 nJ
Neuton
844 nJ
TF Lite
653 nJ

Gearbox · Industrial condition monitoring

2.6× vs Neuton

FICUS
54.4 nJ
Neuton
142 nJ
TF Lite
590 nJ

Transport · Asset tracking and logistics

20.3× vs Neuton

FICUS
28.8 nJ
Neuton
586 nJ
TF Lite
1,275 nJ
Full results tableEvery metric · historical accuracy

Bearing

Bearing: accuracy, energy and time per decision by engine
EngineAcc %nJµs
FICUS100.0011.001.336
Neuton99.84456.7758.079
TF Lite88.62850.13100.351

Toothbrush

Toothbrush: accuracy, energy and time per decision by engine
EngineAcc %nJµs
FICUS100.004.410.539
Neuton100.00640.9879.839
TF Lite24.00923.45108.789

Magic Wand

Magic Wand: accuracy, energy and time per decision by engine
EngineAcc %nJµs
FICUS99.8412.851.572
Neuton99.19843.72104.770
TF Lite81.95653.0475.274

Gearbox

Gearbox: accuracy, energy and time per decision by engine
EngineAcc %nJµs
FICUS98.8854.446.648
Neuton98.48141.6017.788
TF Lite95.75590.0668.045

Transport

Transport: accuracy, energy and time per decision by engine
EngineAcc %nJµs
FICUS100.0028.823.326
Neuton95.09585.7775.431
TF Lite79.751,274.70152.135
Boundary, formulas and C0 energy
  1. Sensor
  2. Acquisitionoutside
  3. Feature extraction
  4. Features in RAMC1
  5. Preparation
  6. Reflex classifierC0 + C1
  7. Class

C1 · processed features → class · headline

E [nJ] = 3.0 V × Q [nC]

P [mW] = E [nJ] ÷ t [µs]

E_corrected [nJ] =

E_gross [nJ] − P_idle [mW] × t [µs]

P_idle is the measured idle baseline; t is the inference window.

C0 · classifier alone, gross energy per decision
DatasetFICUS (nJ)Neuton (nJ)TensorFlow Lite (nJ)
Bearing6.36308.7745.6
Toothbrush2.79494.0803.6
Magic Wand7.39774.7610.8
Gearbox40.25110.2552.2
Transport27.66not measurednot measured
Memory detailsComponents only
FICUS linked code and data, and persistent RAM, per Reflex
ReflexLinked code + data (B)Persistent RAM (B)
Bearing16,9360
Toothbrush7,6640
Magic Wand32,3920
Gearbox4,6560
Transport8,684408 per stream
Instinct memory →

Measured · historical accuracy · 4 Nov 202623456

FIG. E4Study 2 · eye tracking

Study 2 · eye tracking

402 nJ per pupil estimate.

Two Reflex software profiles trade speed for precision.

41–161×

faster · energy-first

29–136×

less energy · energy-first

25–45%

lower mean error · energy-first

Lowest energy and fastest: 402 nJ and 39.46 µs per X/Y update. Lower mean and median error than all three references; P95 error slightly higher than Axon and TensorFlow Lite.

FICUS Energy-first: how much faster, how much less energy and how much lower the error is than Neuton, Axon and TensorFlow Lite. Cells marked worse are FICUS losses.
FICUS vsNeutonNordicCPUAxonNordicNPUTF LiteOpen sourceCPU
Faster per X/YFICUS 39.46 µs41.23 times faster40.60 times faster160.98 times faster
Less energy per X/YFICUS 402 nJ96.57% less energy, 29.12 times less96.83% less energy, 31.58 times less99.26% less energy, 135.89 times less
Lower P95 errorFICUS 61.82 px22.72% lower P95 error5.67% higher P95 error, FICUS is worse5.76% higher P95 error, FICUS is worse

Published factors · computed percentages · historical accuracy

Watch the replay →
All measurementsMeasured · historical accuracy

Lowest energy and fastest: 402 nJ and 39.46 µs per X/Y update. Lower mean and median error than all three references; P95 error slightly higher than Axon and TensorFlow Lite.

FICUS Energy-first: how much faster, how much less energy and how much lower the error is than Neuton, Axon and TensorFlow Lite. Cells marked worse are FICUS losses.
FICUS vsNeutonNordicCPUAxonNordicNPUTF LiteOpen sourceCPU
Faster per X/YFICUS 39.46 µs41.23 times faster40.60 times faster160.98 times faster
Less energy per X/YFICUS 402 nJ96.57% less energy, 29.12 times less96.83% less energy, 31.58 times less99.26% less energy, 135.89 times less
Lower mean errorFICUS 21.40 px45.19% lower mean error25.97% lower mean error25.32% lower mean error
Lower median errorFICUS 15.71 px55.76% lower median error40.63% lower median error39.70% lower median error
Lower P95 errorFICUS 61.82 px22.72% lower P95 error5.67% higher P95 error, FICUS is worse5.76% higher P95 error, FICUS is worse

Each cell is the FICUS margin over that model; ember, marked worse, where FICUS loses. Speedups are the factors published with the measurements; energy and error percentages are computed from the measured table. Accuracy is descriptive: the reused INI-30 validation cohort.

Eye tracking: the complete measured comparison of four FICUS builds, Neuton, Axon and TensorFlow Lite
ModelRuns onTime per X/Y(µs)Energy per X/Y(µJ)Active power(mW)Mean error(px)Median(px)P95 error(px)
FICUS energy-firstReflex software · SR9CPU only39.4610.402410.19921.39615.71461.817
FICUS accuracy-firstReflex software · TDKCPU only89.6410.915910.21718.42112.89453.944
FICUS previous referenceReflex software · KPrevious reference buildCPU only53.1770.52909.94824.36217.95469.389
FICUS alternativeReflex software · TR5Not featured: its historical TEST performance missed the predefined toleranceCPU only77.2460.792010.25318.33913.07053.457
NeutonNordic · Neuton X/YCPU only1,626.81211.71987.20439.03635.51879.996
Axon CNNNordic · Conv/Axon NPUCPU + Axon NPU1,601.91812.70767.93328.90026.46858.500
TensorFlow Lite CNNOpen-source framework · Conv CNN, INT8CPU only6,352.28454.68838.60928.65026.05858.451
Projected at 120 updates/s9Projection · zero-idle average

Projected at 120 updates/s · one 8.3 ms frame each

  • FICUS energy-firstCPU only0.048 mW
    busy 0.5% of each frame
  • FICUS accuracy-firstCPU only0.110 mW
    busy 1.1% of each frame
  • NeutonCPU only1.406 mW
    busy 20% of each frame
  • Axon CNNCPU + Axon NPU1.525 mW
    busy 19% of each frame
  • TensorFlow Lite CNNCPU only6.563 mW
    busy 76% of each frame

Zero-idle average = measured energy per X/Y × 120, the update rate of Ganzin AURORA IIE-class eye-tracking silicon. The Cortex-M33 keeps its 128 MHz clock, which sets how fast each update finishes. Active power (10.199 mW for energy-first) is power while an update executes; the projection averages over a whole second, a different operating condition. These are arithmetic projections: measured energy per update × 120, as if nothing drew power between updates. Board idle power, wake, sensor and illumination are excluded; this is not a measurement of a complete 120 Hz product.

Measured · nRF54LM20B · Nov 202678

FIG. E5Study 3 · MNIST

Study 3 · accuracy-matched MNIST

41 nJ per digit.

lite12 meets Neuton’s accuracy level, P3 fast meets Axon, TensorFlow Lite and CAIR Tsetlin Machine.

148×

less energy than Neuton · lite12

157×

faster than Neuton · lite12

+1.94pp

accuracy vs Neuton · lite12

Each reference on the nRF54LM20B against the FICUS build at its accuracy level: how much faster, how much less energy, accuracy difference
ReferenceFICUS buildFasterLess energyAccuracy gain
Neutonlite1294.76%computed157.27×148.29×+1.94 pts
Axon CNN · NPUP3 fast99.18%published43.72×43.89×+0.24 pts
TensorFlow Lite CNNP3 fast99.18%published635.60×555.28×+0.24 pts
CAIR Tsetlin MachineP3 fast99.18%published255.57×204.35×+1.80 pts
NeutonP3 fast99.18%published · reference21.34×17.51×+6.36 pts
All measurementsMeasured · validation accuracy
MNIST on the nRF54LM20B: accuracy on the common 5,000-image validation set, time, energy, active power and model payload per inference
ModelRuns onValidation accuracyTime µsEnergy nJActive mWPayload KiB
FICUS lite12vs NeutonM33 CPU94.76%4.8341.358.5736.00
FICUS lite24M33 CPU97.16%8.1969.068.4372.00
FICUS lite64M33 CPU98.20%11.29110.259.76192.00
FICUS P3 fastvs Axon · TF Lite · TsetlinM33 CPU99.18%35.60350.269.841,303.97
NeutonM33 CPU92.82%759.596,131.648.0715.98
Axon CNNAxon NPU98.94%1,556.4015,373.749.8890.11
TensorFlow Lite CNNM33 CPU98.94%22,624.50194,490.408.6059.85
Tsetlin MachineM33 CPU97.38%9,097.3071,574.377.871,130.61

Measured · nRF54LM20B · Nov 20261011

Historical · host CPU

A digit in 97 ns.

FICUS PURE runs on a desktop core, not a microcontroller.

21,861×

faster than ResNet-6

3,451×

faster than ViT-4

1,422×

faster than CAIR Tsetlin Machine

Host results and work per digitHistorical · computed ratios
Cost metric
Pick a reference row.

What FICUS PURE does per digit · average over the test set

  • 750pixel comparisonsworst case 2,935
  • 1,063graph-index bit testsworst case 4,800
  • 133ten-class score addsworst case 600

No multiply-accumulates and no floating point. ResNet-6 runs 321.6 million MACs per digit and ViT-4 28.9 million. Zero MACs still means padding, gathers, compares and integer adds.

Historical · no energy measured · 12 Nov 20261213

Next step · Reflexes for Things

Measure it on your product.

Your workload, your target, against your current model.

Device makers

Pilot a Reflex

Measure two or three workloads against your current model.

Start a pilot →

Engineers

Try Reflex Lab

Compare measured Reflexes and profile your own CSV.

Open the Lab →

Investors

The full evidence pack

Executive brief, engineering appendix and KPI annex for approved investors.

Request access →

Notes

  1. 1Average: five C1 Reflexes, eye tracking energy-first, MNIST (lite12/Neuton, P3 fast/TensorFlow Lite) · comparator/FICUS time/energy ratios, equal-weight geometric means · nRF54LM20, 128 MHz, 3.0 V · boundaries differ, accuracy never averaged · Axon untested on five Reflexes; host excluded · excluding 24%-accurate TensorFlow Lite toothbrush: 74.6× slower, 72.2× more energy.
  2. 2Study 1: “Ficus — Reflex cards and Instinct — Nordic evidence v6.3”, pp 1–5 · absolute KPI annex, 3.0 V · Nordic executive brief/engineering appendix, 4 Nov 2026 · nRF54LM20B DK, Cortex-M33, 128 MHz, CPU only · Nordic PPK2, about 1 kS/s · 64 phases, 784 windows · 5 tasks, 6,288 rows.
  3. 3C1: processed float features in RAM → preparation → published class · excludes sensing/acquisition/upstream feature extraction/shared Instinct · C0: classifier alone, preparation untimed · gross includes inference-window idle · idle-corrected subtracts measured baseline, FICUS only · power: E ÷ t, repeated calls · four tasks: 16 anchors; Transport: 163 stored rows.
  4. 4Study 1: frozen artifacts, unequal historical accuracy, not unseen-data accuracy/battery life · TensorFlow Lite toothbrush: 24%; excluding it: 36× slower, 37× more energy, not 50.8×/52.4× · Gearbox/Neuton: 2.7× faster, 2.6× less energy · five-task C1 geometric means, equal weights, Instinct excluded · FICUS: 8.2–8.7 mW, 0.54–6.65 µs.
  5. 5Measured charge × assumed constant 3.0 V · declared rail, not independently verified/traced · absolute nanojoules scale with voltage · Transport C0 Neuton/TensorFlow Lite unmeasured in pack · reported precision retained.
  6. 6Study 1 memory: linked code/data, persistent RAM, not totals · inputs/stack/RTOS/drivers/fixtures additional · Neuton excludes runtime/scaling · TensorFlow Lite excludes interpreter/operators, reserves 32 KiB arena · matched total flash/peak SRAM unmeasured.
  7. 7Study 2: FICUS eye-tracking KPI package, Nov 2026 · nRF54LM20B, Cortex-M33 at 128 MHz · CPU: Reflex software/Neuton/TensorFlow Lite; NPU: Nordic Axon · prepared event-count map → X/Y, frontends included · excludes sensor/illumination/raw-event accumulation, not camera-to-gaze power · charge × declared 3.0 V, untraced; power × time = energy · published factors, computed percentages · error: pupil-centre distance, sensor pixels · reused 5,396-row INI-30 validation, different training histories, not prospective.
  8. 8Energy-first SR9: P95 61.82 px, 5.7–5.8% worse than Axon (58.50 px)/TensorFlow Lite (58.45 px) · accuracy-first TDK: lower mean/median/P95 vs all three · FICUS active: 10.2 mW vs 7.2–8.6 mW, finishes 17–161× sooner · K: previous reference · TR5: lower mean/P95 than TDK; historical TEST missed predefined tolerance, unfeatured · SR9/TDK memory/placement, operation counts, quiet-floor projection unavailable.
  9. 9120 updates/s: arithmetic projections, energy/update × 120, hypothetical zero idle · not a complete 120 Hz product measurement · excludes board idle/wake/sensor/illumination · active 10.199 mW vs projected 0.048 mW: different conditions · clock 128 MHz, not 120 Hz · 39.46 µs/update, 0.47% busy/8.3 ms frame · rate: Ganzin AURORA IIE, Electronic Specifier, 4 Jan 2026.
  10. 10Study 3: “FICUS MNIST results on nRF54LM20B”, Nov 2026 · Nordic nRF54LM20B DK, Cortex-M33, 128 MHz, 3.0 V · Nordic Axon CNN: NPU; others CPU · 28 × 28 image → digit · separate firmware, identical images, different memory placements, descriptive factors · common 5,000-image validation, official test separate · lite12 94.76% vs Neuton 92.82%; P3 fast 99.18% vs Axon/TensorFlow Lite 98.94%, CAIR Tsetlin Machine 97.38% · full files: 28 comparisons/historical implementations · MNIST: LeCun, Cortes, Burges.
  11. 11MNIST: P3 factors published, unrounded measurements · lite12/Neuton computed, rounded table, ≈±0.1% rounding · muted P3/Neuton: published reference, not accuracy match · larger tables: lite12 36 KiB/Neuton 16 KiB; P3 1,304 KiB/TensorFlow Lite 60 KiB/Axon 90 KiB · active lite12 8.57 mW/Neuton 8.07 mW; P3 9.84 mW/others 7.87–9.88 mW · saving: time, not power · payload excludes code/runtime/working RAM · peak RAM/operation counts unavailable · full files: above-idle energy, idle power, charge, sensitivity.
  12. 12Host: “Updated MNIST Ficus results”, 12 Nov 2026 · AMD Ryzen 9 9950X, one thread · 16 cores · up to 5.75 GHz · AVX-512 · warm CPU · raw 28 × 28 uint8 pixels in RAM → padding/predicates/decision → class · excludes file I/O/setup/training · B1/B128: 1/128 digits/native chunk · median of five window means, not p50/p99 requests · audit 800/800: source/frozen hashes/arithmetic, no rerun · historical official 10,000 MNIST test digits · weights trained; thresholds from 5,000 validation digits.
  13. 13Desktop, not microcontroller; no energy/power measured · graphs: 12 Nov; references: 10 Nov, same CPU model/different cores · B1: reference 1,024 digits/FICUS 10,000, descriptive ratios · PURE accuracy: −0.46 pp vs ResNet-6, −0.28 pp vs ViT-4, +1.46 pp vs CAIR · local implementations, not all optimised runtimes · zero MACs/float still requires padding/gathers/comparisons/integer adds/memory moves · arrays, not RAM/MCU flash · PURE: new fit, 4.9× raw32 tables, not compression · Tsetlin cached-index workspace ≈502 MB.
  14. 14Historical accuracy: execution, not fresh generalisation · Transport 163/163: 156 known-session rows, 7 observer rows · The Nordic average spans three boundaries · one MCU family · pilots: whole-device sensing/radio/idle energy, battery life vs customer baseline · FICUS Silicon: design target until V5 measured · TensorFlow Lite: open-source framework, now published as LiteRT · Neuton/Axon NPU: Nordic.