Premium AI features in a new form.
Run complex models in a hearing aid, a ring, a patch – no bulky battery, no redesign.
GPX10PRO
Full AI inference. On a coin cell.
The world's first energy-aware AI processor running real neural networks, not rule-based shortcuts, at microwatt power. Built on the A-Cube architecture.
What GPX10 Pro unlocks
The chip that ends the power-vs-intelligence tradeoff.
For a decade, product makers chose a dumb MCU that lasts months, or a smart NPU that dies by lunch. GPX10 Pro is the first that refuses to choose.
Premium AI features in a new form.
Run complex models in a hearing aid, a ring, a patch – no bulky battery, no redesign.
Months on a coin cell.
Always-on AI at ~80 µW. Ship the battery life your reviews live or die on.
Private by default.
Data never leaves the device. No cloud round-trip, no latency, no privacy liability.
One chip replaces the stack.
MCU + AI accelerator + sensor hub + memory you're juggling today and it stays aware while it sleeps.
The signature mode
Always on. Never asleep.
GPX10 Pro runs AI around the clock at microwatts, and the instant something matters, it surges to full power. No reset. No waking up. It was never off.
AI CORE
HOST CPU
power
< 100 uW
always on AI
Continuous pulse
Performance
512 GOPS
instant
REFLEX SURGE
Subconscious
Reflex Surge
Return to the baseline
Measured proof
Not projected.Measured in silicon.
Here's GPX10 Pro against the alternatives a design team actually weighs.
GPX10 Pro
Peak Compute (GOPS)
512
Lowest Active AI power
200 µW
Measured AI Efficiency
7.3
Typical MCU
Peak Compute (GOPS)
0.02
Lowest Active AI power
20 mW
Measured AI Efficiency
0.02
MCU + NPU
Peak Compute (GOPS)
100
Lowest Active AI power
1 mW
Measured AI Efficiency
1.2
Proven in production
The empirical advantage.
Real products, real numbers, with the AI running the whole time.
The Safety Win
90 Days
on a single charge
A pendant that watches over you, running fall detection and SOS voice recognition continuously, on-device, for 90 days on a single charge. No phone to carry, no cloud to depend on. Protection that never goes dark.
Coming Soon
The Women’s-Health Win
>98%
fall & assault recognition
In DimensionNXG’s MAI wearable, GPX10 Pro recognizes falls and assault patterns on-device at >98% accuracy and still runs 14 days on a 200 mAh cell. Detection stays local and private; the radio fires only when there’s a real event to report.
The Industrial Win
2+ Years
on one cell
Predictive maintenance that lives on the machine, continuously analyzing vibration and sound to catch anomalies before it breaks, entirely at the edge. A retrofittable, battery-powered node that runs 2+ years on a single cell, sending alerts, not raw data.
Coming Soon
BEST-FIT APPLICATIONS
Built for always-on.
Proven across markets.
The same chip, tuned for the job, from a wrist to a factory floor.
HEARABLES
Always-On Wake Words
Run continuous wake-word detection and voice commands locally inside earbuds without draining the tiny audio battery.
Adaptive Voice Isolation
Separate the user’s voice from noisy surroundings in real time, directly on-device, without relying on phone or cloud processing.
Ready to Run
GPX10 Pro already speaks your application.
Every application above runs on a real, GPX-tuned model from the Model Zoo, open-source and Ambient-built, deployable in one click. Flash one to your kit and watch it run in seconds.
Build with GPX10PRO
Your models. Your IDE. No rewrites.
A new architecture shouldn't mean a new way of working. With ModelForge, it doesn't.
TRAIN
Bring your TensorFlow, Keras, or ONNX model. Or start from our pre-trained library.
Browse the Model ZooCOMPILE
Push-button: ModelForge quantizes and maps it onto A-Cube. No manual translation.
DEPLOY
One unified build in standard Eclipse. Up to 90% of your existing C code ports over.
From eval to production
From bench to volumewithout rewriting a thing.
The C code, the build, the AI you validate on the kit ports straight to production silicon. This is the part competitors can't offer.
Evaluate
Nurasense EVK
A complete dev kit with onboard sensors, camera, mics, and pre-loaded demos. Measure the power yourself, day one.
Integrate
SOMs module
Drop our pre-engineered System-on-Module into your carrier board. Skip the RF, power, and sensor-routing nightmare.
Scale
GPX10 Pro Silicon
The raw SoC for high-volume production.
Architecture
Everything in one chip.Nothing wasted.
A complete edge-AI SoC AI engine, control, sensing, memory, and security integrated so your board doesn't have to be.
A-Cube compute
10 MX8 cores, 2,560 MACs/cycle, replaces a separate AI accelerator.
Two power domains
A 5-core island sips microwatts; the rest powers down.
Integrated sensing
Up to 10 sensor streams fused on-chip; no external sensor hub.
Specs & resources
The full picture.
Control
+ARM Cortex-M4F (32-bit, FPU)
Sensing & Analog
+14-bit SAR_ADC
+PGA
+Sensor-fusion DMA
+Low Battery Indication
Temperature
+-40 to +85C
Package
+128BGA Package
+8mmx8mm
+CSP 3.2x3.2 mm On demand
Security
+128 Bit KEY AES-DES
Peripherals
+QPI
+SPI
+I2C Master
+I2C Slave
+UART
+GPIO
+I2S Master
+DVP
+ADC
Memory
+120KB L0 RAM for input operands and coefficients for AI Processor cores
+2MB unified L1 SRAM
+64KB retention RAM
+32KB Video buffer
+16KB multi bank buffer for sensor fusion
Power
+Core 1.2V (0.9V - 1.3V) + Analog I/O 3.3V (1.8V - 3.3V)
+Two Power Domains (Always On and Non Always On)
AI Cores
+5x Always ON AI cores
+5x AI Cores in Non Always ON domain for additional workload
+Dedicated Preprocessing stages - FFT (256 point), Square, Shift and log
Start building with GPX10 Pro.
A complete Edge AI SoC integrates the AI engine, control, sensing, memory, and security, so your board doesn't have to.
Get an Evaluation Kit.
Explore how Ambient AI can unlock new capabilities in your wearable product. Strategic planning session with our applications team.
Request Eval KitScale to increase the volume.
Be the first to access our upcoming Vision, Sound, and Industrial modules.
Talk to Sales