GrAI One, the first instantiation of the GrAI Matter's NeuronFlow technology, is implemented in the 28nm HPC manufacturing process from TSMC.
The NeuronFlow architecture combines ideas of dataflow processing with neuromorphic computing. GrAI Matter claims that GrAI One has low latency making it suitable for applications in autonomous navigation, human-machine interaction and smart healthcare markets.
The die is 20 square millimeters in area and implements a mesh of 196 neuron cores with local neuron/synapse memories. The neuronal cores process either 8bit or 16bit data and with each core representing 1,024 neurons the overall chip can scale to represent up to 200,000 neurons. The chip is essentially a fabric of neuron cores connected by packet-switched network-on-chip.
GrAI One block diagram: Weights are stored in local SRAM and can be shared. Data and weights are passed to neurons that have state in local SRAM and perform basic neural and ALU functions. Source: GrAI Matter Labs.
The chip provides a GPIO interface to offload latency-critical AI workloads of a host processor. At 100 percent neuron core utilization, GrAI One may consume as little as 35mW at 25 MHz clock frequency.
Next: Presented to Linley Gwennap