NimbleAI develops a neuromorphic 3D‑stacked chip that combines light‑field and depth sensing with event‑driven neural processing, mimicking retinal and insect eye principles to handle only salient visual changes. By fusing sensor, memory, and compute in a single silicon volume, the chip delivers orders‑of‑magnitude improvements in energy efficiency, latency, and area over traditional CPU/GPU video pipelines, enabling ultra‑low‑power edge AI vision for robotics, autonomous vehicles, and semiconductor manufacturers.
Funding
Funding not disclosed
Founders
Product
Problem
Current visual processing systems rely on frame-based video pipelines executed on CPUs or GPUs, which consume high power, exhibit latency, and require large silicon area, limiting their suitability for edge AI and low‑power applications.
Solution
NimbleAI is developing a neuromorphic sensing‑processing chip that integrates light and depth sensing with event‑driven neural computation in a 3D‑stacked silicon volume. The chip emulates retinal light detection by responding only to changes in illumination and adopts insect‑inspired compound‑eye depth perception, reducing redundant data. Processing follows brain‑like principles: only significant neuron state changes are propagated, and adaptive visual pathways adjust temporal and spatial resolution at runtime. By fusing sensing, memory, and compute in a single 3D stack, the architecture achieves orders‑of‑magnitude improvements in energy efficiency, latency, and area compared with conventional CPU/GPU video pipelines, enabling advanced AI and computer‑vision algorithms on ultra‑low‑power platforms.
Target Audience
Primary customers are edge‑AI hardware manufacturers, autonomous‑vehicle and robotics system integrators, and semiconductor companies seeking ultra‑low‑power vision solutions.
Features
- Event‑driven light‑field dynamic vision sensor that captures monocular depth without frame buffering
- Neuromorphic processing core that propagates only salient neuron spikes, minimizing unnecessary computation
- Adaptive visual pathways that reconfigure sensing and processing resolution in real time
- 3D‑stacked silicon integration of sensor, memory, and compute to maximize bandwidth and reduce footprint (~50 mm²)
- Compatibility with standard convolutional neural networks, delivering a full perception stack on‑chip
- Prototype development platform with programming tools and EDA support for 3D integration