Skynet

Skynet is a framework architecture for efficiently simulating large feed-forward neural networks in a distributed fashion.

Distributed Simulation

Skynet exploits the parallelism inherent in the nature of perceptrons by implementing pipelined simulation of neurons. The resultant architecture resembles the map-reduce algorithm.

Linear Speedup

The time needed to simulate a test halves every time the number of cores are doubled.

Shown here, the simulation times for the number of threads spawned on a logarithmic scale. The test network comprised of 11,100 nodes and 11 million synapses for 1000 input vectors, and the simulation was run on a 24-core machine.

Super-Linear Scalability

As the number of input vectors is increased, the simulation time scales up super-linearly.

Shown here, the simulation times for the number of input vectors on a logarithmic scale. The test network comprised of 11,100 nodes and 11 million synapses for 16 cores, and the simulation was run on a 24-core machine.

Get it now.

Code

Paper