Weak scaling. If, instead of dividing the execution times by the total number of synapses, we divided it by the number of synapses assigned to each computational core, we should obtain, for the same consideration of the previous section, a value that for an ideal code, executing on an ideal machine should be constant for different network sizes and number of computational cores assigned to the solution of the problem. Figure 3-5 is the graph of our measures. Looking for hints about how to further enhance the scalability features of the system, we observe an interesting feature: when the simulation is distributed among a large number of hardware cores, the simulation runs relatively faster for larger configurations (i.e. configurations that host more “neural columns” per hardware core/software process). We analyze this point in the next “Discussion” section. Figure 3-5 Weak scaling of the DPSNN-STDP benchmark
