In the example of Figure 2, the first non-zero is a2 on PE2. The value a2 and its column index 2 is broadcast to all PEs. Each PE then multiplies a2 by every nonzero in its portion of column 2. PE0 multiplies a2 by W0,2 and W12,2; PE1 has all zeros in column 2 and so performs no multiplications; PE2 multiplies a2 by W2,2 and W14,2, and so on. The result of each product is summed into the corresponding row accumulator. For example PE0 computes b0 = b0 + W0,2a2 and b12 = b12 + W12,2a2. The accumulators are initialized to zero before each layer computation.