Follow Us On

End-to-End fully Analog, a quantum leap in AI Training efficiency, and an end to GPU Accelerators

September 22, 202003:00pm - 04:00pm

Speaker

Gordon Wilson

Gordon Wilson

  • September 22, 2020
  • 03:00 pm - 04:00 pm

Abstract

Rain has created analog neurons that are tiled across the chip surface creating a high-density CMOS neuromorphic processor. Instantiating millions of artificial neurons in a square centimeter enables massive neural networks to fit onto a single chip. The memristor is recognized as an ideal artificial synapse. Memristor is a combination of "memory" and "resistor"; it is a resistor capable of changing its resistance. Rain’s memristive synapses exist in the shells of the interconnect wires, stacked vertically on top of the neurons. The nanowires are randomly arranged on the surface of the chip. Emerging from this random placement is a powerful, sparse, brain-like topology: a small-world network. This connectivity allows the network to remain well connected even at massive scale. Using high efficient analog processing allows the device to provide GPU level of AI capabilities at a fraction of the operating cost.