The AI engine on the GPX10 Pro can support all important neural networking model types including CNNS, RNNS, LSTMS and Grus. Local at the edge, Ambient says the device offers up to 100 times more power, performance and area than traditional 32-bit microcontrollers.
Based on Ambient Scientific's proprietary Digan Silicon Architecture, the GPX10 Pro can map the matrix multi-pre-operation and activation flow of neural network models directly to in-memory analog computer blocks, a structure that removes the wasteful cycles and overhead of the typical pulley instruction set of traditional processors.
As a result, the GPX10 Pro can perform common edge AI functions such as speech recognition, keyword spotting, low-frequency computer vision, and intelligent sensing faster, and can use far less power than MCU, NPU, or GPU.
“Today's MCUs and NPUs are hit by traditional silicon architectures when they try to run AI models. It's like hitting baseball with a tennis racket. This is the wrong tool for work.” “The GPX10 Pro shows what is possible when building an architecture natively for AI. This is 100 pieces of AI performance at microwatts of power.”
The GPX10 Pro is a highly integrated SOC that allows for local AI inference, but can also be used on devices with a single coin cell battery.
AI processing runs on two sets of five MX8 AI cores in two separate power domains. One set is always on-block to support interfaces and fusion of ultra-low force power sensors when performing constant-on keyword spotting, when the chip performs consumption of less than 100µW. The 10 MX8 core performs up to 2,560 multiplication accumulation (MAC) operations per cycle, generating a total peak AI throughput of 512 GOPs.
The GPX10 Pro's computational functions are supported by 2MB of on-chip SRAM to allow for the implementation of larger and more complex AI models.
The GPX10 Pro also features an ARM Cortex-M4F CPU core for control functions. Integrated analog features include ultra-low power ADCs, enhanced I2S logic, and interfaces with up to 8 simultaneous analogues and 20 digital sensors.
Additionally, Ambient Scientific offers a nebula AI enablement toolchain to accelerate the training, development and deployment of AI models for GPX10 and GPX10 Pro. Compatible with other model training frameworks such as Tensorflow, Keras, ONNX, and more.
The AI core of the chip programmable with the Nebula toolchain provides designers with the flexibility to adapt to the types and topology of AI models that are evolving.
Ambient Scientific also offers a SenseMesh hardware sensor fusion layer. This allows for low-latency sensor fusion by connecting multiple sensors to the core via tightly coupled meshes. This generates an immediate response to trigger events and ultra-low idle mode power to offload sensor polling from the CPU.
The GPX10 Pro is currently available for sampling. Volume production is expected to begin in the first quarter of 2026.
