Ambient Scientific has introduced the GPX10 Pro, a next-generation system-on-chip (SoC) designed to bring powerful, efficient AI processing to ultra-low-power edge applications. Built on the company’s proprietary DigAn® architecture, the chip enables local AI inference on devices as small as a coin cell battery.
The GPX10 Pro is optimized to support a wide range of neural network models—including CNNs, RNNs, LSTMs, and GRUs—directly at the edge. It offers up to 100x improvements in power efficiency, performance, and footprint compared to conventional 32-bit microcontrollers. At the heart of the innovation is Ambient Scientific’s analog-in-memory compute design, which allows matrix-multiply and activation operations to run directly within memory blocks. This eliminates the inefficiencies of traditional instruction-based processing and significantly reduces energy consumption.
With its ability to handle tasks such as keyword spotting, voice recognition, low-resolution computer vision, and smart sensing, the GPX10 Pro operates faster and more efficiently than typical MCUs, NPUs, or GPUs. “Trying to run AI on conventional architectures is like hitting a baseball with a tennis racket—it just doesn’t work well,” said GP Singh, CEO of Ambient Scientific. “The GPX10 Pro proves what’s possible when you build your architecture from the ground up for AI—hundreds of GOPs at microwatt power levels.”
The chip integrates two clusters of five MX8 AI cores, separated into distinct power domains. One operates in always-on mode, consuming under 100µW during continuous tasks like keyword spotting. In total, the ten cores can achieve 512 GOPs of AI throughput, handling up to 2,560 MAC operations per cycle.
To support larger AI models, the chip includes 2MB of on-chip SRAM—10x more than its predecessor, the GPX10. The GPX10 Pro includes a range of built-in features to support both AI and general-purpose tasks. It integrates an Arm Cortex-M4F processor for traditional control functions, along with an ultra-low-power analog-to-digital converter (ADC) and advanced I2S logic for audio and data processing. The chip also provides interfaces for up to 8 analog and 20 digital sensors, making it ideal for complex, sensor-rich edge applications.
To accelerate development, Ambient Scientific offers a complete software toolchain. The Nebula™ platform supports major AI frameworks including TensorFlow, Keras, and ONNX, enabling seamless training and deployment of models on both the GPX10 and GPX10 Pro. Additionally, the SenseMesh™ hardware layer allows multiple sensors to be connected through a low-latency mesh network, offloading sensor polling from the CPU to reduce power consumption and improve system responsiveness.
Ambient Scientific is demonstrating real-world applications powered by the GPX10 Pro at Electronica 2025 in Bangalore, India. Visitors to Booth 5.E31 (September 17–19) can see live demos of fall detection, voice recognition, and face identification, all running efficiently on coin cell-powered devices. The GPX10 Pro is available for sampling now, with mass production planned for Q1 2026.
Ambient Scientific is a fabless semiconductor company specializing in energy-efficient AI processors built on its DigAn® architecture, which blends the efficiency of analog with the scalability of digital. Its product ecosystem—including CubicCore™ compute, SenseMesh™ fusion, and Nebula™ software—redefines how hardware and software co-evolve to enable energy-aware intelligence across edge-to-cloud devices. Headquartered in Santa Clara, the company is committed to delivering sustainable AI solutions for everything from compact sensors to massive data centers.
