New semiconductors that merge compute and vision
TLDR: Synapse Semiconductor builds image sensors that process AI on the sensor itself. Each pixel captures light and runs neural-network computation in the same device, replacing the need for a GPU while cutting the power and latency of edge vision.
A camera has no use by itself. Its value is only when its information gets evaluated. How can a camera that only has value when combined with a GPU support local intelligence? GPUs will get better, yet the fundamental problem is the separation of the camera and the GPU. Solving that separation will unlock bountiful innovation in machine perception.
RETINA: a chip designed like the human retina which runs neural network operations in the camera pixel.
We have compressed the whole edge AI vision hardware stack into one wafer. CNNS, VLMS, etc can run on just one chip that senses light. Our compute transistors are also the same photosensors that see light. There is no longer separation between compute (Nvidia Jetson Nano) and a camera (RealSense Depth Camera).
If you are building robotic perception, drones, and vision systems for physical AI —> book.synapsesemi.org
