Estimates the time needed to compute forward pass of neural net of size of human brain (10^15 synapses). Assumes that synapse implements function which can be well approximated by multiplication.

brain.py

import numpy as np

import theano
from theano import tensor as T

# synapses to compute at once
N = 2 ** 13
M = N
Sym_rep = 16
x = np.random.randn(N).astype('float32')

print "compiling ..."
xt = theano.shared(x)

yt = xt[:N]

for i in range(Sym_rep):
    Wt = theano.shared(np.random.rand(M, N).astype('float32'))
    yt = T.maximum(0, T.dot(Wt, yt) )

fp = theano.function(inputs=[], outputs=[yt], allow_input_downcast=True)

print "computing ..."

# number of synapses simulated
S = N * M * Sym_rep

# number of synapses in human brain
H = 10 ** 15

from time import time
reps = 10

st = time()

for i in range(reps):
    #y = np.maximum( 0, np.dot(W,x) )
    y = fp()

d = time() - st
perbatch = d / reps
brain_forward = perbatch * (H / S)

print "time in seconds for forward pass of whole human brain, s: ", brain_forward
print "time in seconds for forward pass of whole human brain, h: ", brain_forward / 3600