Ambient Scientific GPX10 Pro MCU delivers years of always-on AI on a coin-cell battery

Ambient Scientific GPX10 Pro is a processor/microcontroller for always-on embedded AI applications on power-constrained edge devices. With a consumption of under <100μW, it can deliver years of always-on AI on one coin-cell battery.

The SoC features ten software-programmable MX8 AI cores leveraging the company’s DigAn technology, ultra-low power ADC, an 8-bit DVP camera interface, 2048 KB SRAM, support for numerous peripherals, and an Arm Cortex-M4F core for non-AI workloads.

GPX10 Pro key features and specifications:

• MCU core – Arm Cortex-M4F core clocked at 100 KHz to 100 MHz
• Memory
  • 2048 KB system SRAM
  • 64KB retention SRAM in “Always On” block
• Storage – External SPI or QSPI flash
• 10x MX8 programmable AI Cores for ultra-low power deep learning applications
  • Based on the DigAn AI engine, which can change behavior between being a master and a slave.
  • Up to 2,560 MAC operations per clock cycle (256 MAC per cycle per core)
  • Peak performance of 512 GOPS @ 100 Mhz with efficiency > 7 TOPs/W
  • 256-point FFT Engine in “Always On”
  • Support for multiple operand and weight resolutions (4 to 32-bit)
  • Support for standard neural networks (CNN, RNN, LSTM, FCN) as well as custom networks from scratch
• Camera interface – 8-Bit DVP interface and 32KB auto streaming video buffer for low-frequency image classification, I2C for configuration
• Sensor fusion – Connect up to 10 analog and digital sensors simultaneously
• I/Os
  • 16x GPIO, 4x interrupts
  • 1x I2C slave, 1x I2C master, 2x SPI, 2x UART
  • Multi-channel ultra-low power 16-bit ADC (14-bit ENOB) to support up to 8x analog sensors at <5µW at 20K samples/s, <20µW at 1M samples/s
  • Up to 4x analog microphones and 2x digital I2S microphones (16-bit)
• Security – AES-128 Asymmetric encryption
• Low-power design
  • Sophisticated clocking structure (LPO, External Clock Source, PLL) for SoC-level power minimization
  • <100μW power usage
  • Years of always-on AI on one coin-cell
  • Charger-less AI applications are possible by harvesting energy from various natural sources, such as kinetic energy and RF energy

The company provides an SDK offering the following:

• Supported AI/ML Frameworks – Keras, Tensorflow, ONNX
• Compiler Toolchains – ONNX Runtime, MX8 Runtime, ONNX-MLIR
• Ambient Libraries
  • MX8 Math Libraries (Arithmetic Logic, Compare, etc.)
  • MX8 AI Libraries (Convolution, Recurrent, etc.)
• System Software – Peripheral drivers, system interface, emulation support

The company lists “Windows” and “RTOS” for the supported operating systems. That means the Ambient SDK runs on Windows (and Linux, based on the doc I’ve received), while the processor runs an RTOS on the Cortex-M4F core. The GPX10 Pro Development Kit is also available with an analog and digital microphone, an accelerometer, a gas sensor, an integrated camera connector, a Bluetooth module for communication, a serial flash for storage, several I/Os, and JTAG and USB UART for debugging.

Ambient Scientific also releases some performance numbers for Sensor Fusion AI and Vision AI workloads, and compares the 512 GOPS GPX10 to a specialized ASIC chip and a typical MCU with a built-in NPU. Note that’s for the earlier GPX10, and not the new GPX10 Pro, but the ten MX8 cores appear to be the same, and the Pro model mostly features extra memory (2048 KB vs 256 KB) and various enhancements.

There’s limited public information at this stage, and you’d need to request the SDK on the product page to access it. The development kit appears to be available, but the company was not ready to disclose pricing information publicly.

Jean-Luc Aufranc (CNXSoft)

Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.

Share this:

Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress. We also use affiliate links in articles to earn commissions if you make a purchase after clicking on those links.