Page 2 Massive parallel processing demonstrated in a...

Page 2
Massive parallel processing demonstrated in a hardware-only design
In Milestone 1 (announced 13 May 2015) BrainChip demonstrated a software simulation of SNAP versus a traditional genetic algorithm model illustrated by each seeking to navigate a racetrack. The genetic algorithm model, as used by global technology companies, took 15 minutes to complete one circuit whereas the SNAP car completed one circuit of the track in 25 seconds. In Milestone 2 SNAP completes the same circuit in just 0.89 seconds.
SNAP performed 3,571,430 neural update cycles during the 0.89 seconds.
This demonstrates an extremely fast cycle time, and further shows the unparalleled processing power that SNAP can achieve.
Due to the massive parallel processing capabilities of SNAP, it is anticipated that a network of 18,225 neurons will execute 21,578 times faster than a CPU (Central Processing Unit).
Neuron count versus neuron complexity
Neuron numbers alone are meaningless without considering the features of the neuron and its synapses. The processing power of one hardware-only BrainChip neuron is significantly higher than of sigmoid neurons. One external source estimated that a single spiking neuron performs the same task as 4997 sigmoid neurons.*
FIGURE 1: COMPONENTS OF BRAINCHIP NEURONS THAT ENABLE AUTONOMOUS LEARNING.
The illustration above serves to demonstrate the complex nature of the BrainChip neurons versus the simpler sigmoid model.