Measured on Nordic
6
Same-board runtime evidence.12
Method and results →Applications
Pick a device. See its Reflex software, numbers and evidence status.1
Select a device · scroll for more
FICUS · Eye-tracking Reflex
402 nJ per X/Y
● Sensors · Event camera at the nose pad, facing the eye · 640×480
Measured · nRF54LM20
Wearables
Where is the pupil, right now?
Reflex software finds pupil X/Y on the microcontroller.2
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 vs | NeutonNordicCPU | AxonNordicNPU | TF LiteOpen sourceCPU |
|---|---|---|---|
| Faster per X/YFICUS 39.46 µs | 41.23 times faster | 40.60 times faster | 160.98 times faster |
| Less energy per X/YFICUS 402 nJ | 96.57% less energy, 29.12 times less | 96.83% less energy, 31.58 times less | 99.26% less energy, 135.89 times less |
| Lower mean errorFICUS 21.40 px | 45.19% lower mean error | 25.97% lower mean error | 25.32% lower mean error |
| Lower median errorFICUS 15.71 px | 55.76% lower median error | 40.63% lower median error | 39.70% lower median error |
| Lower P95 errorFICUS 61.82 px | 22.72% lower P95 error | 5.67% higher P95 error, FICUS is worse | 5.76% higher P95 error, FICUS is worse |
Each cell is the FICUS margin over that model; ember, marked worse, where FICUS loses. Speedups are the published factors; percentages are computed from the measured table.
Measured · nRF54LM20B · Nov 20262
Illustrations · not product photographs9
Eye tracking, replayed
Two FICUS profiles. Neuton. Axon NPU. TensorFlow Lite.5


Slowed · 4 observations/s · not runtime footage
The ring is the annotated pupil centre, the cross is each model's estimate, and the number under a panel is that frame's error in pixels. The energy and time in each header are the measured mean per update on the board, not per frame.
Measured · nRF54LM20B · Nov 20262


Entertainment
Exploring · same near-eye inputWhy render detail you cannot see?
Detail, menus and aiming follow gaze.8


Manufacturing
Exploring · same near-eye inputNext step, with both hands busy?
Gaze advances steps and confirms inspection.8


Construction
Exploring · same near-eye inputPlans and help, with gloves on?
A glance opens the hard-hat visor.8


Medical
Exploring · same near-eye inputWhat can eye movements reveal?
Track movements for clinical research.8


Research
Exploring · same near-eye inputWhere do shoppers, drivers and athletes look?
Map attention beyond the lab.8


Automotive
Exploring · different opticsEyes off the road?
Gaze can trigger a distraction warning.8
Illustrations · not product photographs8
Published pain benchmarks. Not FICUS results or customer losses.9
€78,068
per fracture, first-year care14
US$2.3M
per hour, large car plant15
US$22,434
per person over 10 years35
Norway
▲ On the charger
€78,068 per fracture, first-year care14
Mean hip-fracture care cost, standardised pathway
Rayong, Thailand
▲ Asleep between checks
US$2.3M per hour, large car plant15
Unplanned downtime cost of a large automotive plant
Jurong Island, Singapore
▲ Nobody watching
~US$500k per hour, large oil & gas site16
Downtime cost at large oil and gas facilities
Pilbara, Australia
▲ Nobody watching
A$25,000 per hour of stopped belt17
Iron-ore conveyor downtime cost
Kazakhstan
▲ Asleep between checks
US$600M for one derailment settlement18
East Palestine derailment class-action settlement
Hanoi, Vietnam
▲ Cries wolf
~US$19B per year in Vietnam19
Economic cost of road deaths and serious injuries
Rotterdam, Netherlands
▲ Asleep between checks
US$10–12M per crane, up to 2-year lead time20
Procurement cost of a ship-to-shore crane
Catania, Italy
▲ Nobody watching
US$126,000 per vacuum-pump incident21
Chamber pump failure with a 24-hour outage
Kigali, Rwanda
▲ Cries wolf
US$66,660 per closed alarm-response claim22
Mean indemnity for unanswered equipment warning alarms
Mexico City, Mexico
▲ Nobody watching
US$39B per year worldwide23
Estimated annual cost/value of non-revenue water
New South Wales, Australia
▲ Too late
US$1.925B in accrued wildfire claims24
PG&E's liability for the 2021 Dixie Fire
Shenzhen, China
▲ Nobody watching
US$258.3M NYC losses in 202325
Upper estimate of e-mobility battery fire losses
Johor, Malaysia
▲ Nobody watching
>US$1M for 1 in 5 major outages26
Share of recent major outages that cost over US$1M
Ouarzazate, Morocco
▲ Too late
US$58.4M per solar hail claim27
Average insured solar hail loss claim
Kolwezi, DR Congo
▲ Too late
US$150,000 per hour, large copper mill28
Contained-metal value a large copper SAG mill processes per hour
Taoyuan, Taiwan
▲ Too late
US$1M per server shipment29
Value of one high-value server shipment
N1 highway, South Africa
▲ Nobody watching
US$46,765 per non-injury truck crash30
Estimated economic cost of a non-injury large-truck crash
Quebec, Canada
▲ Nobody watching
>US$30,000 per freeze-damage claim31
Average insurer payment for freeze-related water damage
West Java, Indonesia
▲ Asleep between checks
US$59M per heavy-industry plant, per year32
Annual unplanned downtime cost of a large heavy-industry plant
Manila, Philippines
▲ Asleep between checks
≈US$5M to replace one substation33
Estimated replacement cost after a substation fire
São Paulo, Brazil
▲ Nobody watching
US$6,386 per AF-related stroke admission34
Median hospital cost of an AF-related ischemic stroke
In conversation
▲ Left to the user
US$22,434 per person over 10 years35
Excess healthcare cost linked to untreated hearing loss
Changwon, South Korea
▲ Too late
US$150,000 per hour, top end for SMEs36
Downtime cost for small and medium manufacturers
In the factory
▲ Asleep between checks
US$253M per large plant, per year37
Annual unplanned downtime cost of an average large plant
Rivers State, Nigeria
▲ Nobody watching
~US$3M revenue lost in two days38
Revenue lost during two days of evacuation downtime
Hamburg, Germany
▲ Nobody watching
~US$35B per year, biopharma39
Losses from failures in temperature-controlled logistics
Madrid, Spain
▲ Nobody watching
>US$361B global occupational heat loss40
Heat-related deaths, incapacity and treatment costs
Mumbai, India
▲ Cries wolf
US$8,761 per range/cooktop fire41
Average home range/cooktop fire property loss
Norway
▲ Easily fooled
€20,700 per stolen insured car42
Average cost to motor insurers of a car theft
Lagos, Nigeria
▲ Nobody watching
US$1.41B per year in US construction43
Struck-by injury workers' compensation costs
North Sea
▲ Needs a crew
US$100k–1M+ per offshore blade repair44
Offshore blade repair once damage has escalated
Atacama, Chile
▲ Too late
US$1.54M per work-related death45
Average economic cost of a fatal workplace injury
Kumasi, Ghana
▲ Too late
US$48,000 per medically consulted injury46
Economic cost of a medically consulted work injury
Cairo, Egypt
▲ Nobody watching
US$6,884 per child asthma admission47
Mean hospital service delivery cost of an asthma admission
Europe
▲ Asleep between checks
£1,172.20 per van-day off the road48
Average UK business cost of van downtime
Malawi
▲ Too late
A$2,300 per cold-chain breach49
Average vaccine value lost in a cold-chain breach
AI-generated scenes · public benchmarks9
Measured on Nordic
6
Same-board runtime evidence.12
Method and results →Historical study
4
Earlier studies. Separate setups.4
Six domains, one compiler →Pilot workload
3
Built from your model. Not measured yet.3
Battery calculator →Next step · Reflexes for Things
We compile Reflex software, verify it, then measure on your target.
Device makers
Pick two or three workloads. Measure Instinct against your current model.
Start a pilot →Engineers
Explore Reflexes, compose Instinct, profile your CSV in the browser.
Open the Lab →