Simulating Neural Network Processors

Table 3

Simulation results of different configurations.

Configuration	Fully connected	Conv2d	Max pooling
	#Cycles MAC utilization (%)	#Cycles MAC utilization (%)	#Cycles MAC utilization (%)
Original	285,379; 91.9	315,290; 93.5	42,779; 15
Larger MAC	266,491; 12.3	292,750; 12.6	42,779; 3.7
Wider scratchpad bandwidth	221,808; 14.8	247,330; 14.9	38,325; 4.2
Larger accumulation buffer	115,187; 28.5	127,809; 28.8	36,171; 4.4
Double bandwidth from off-chip to on-chip	61,923; 53.0	68,417; 53.9	20,771; 7.7