Fig 5. Benchmarks with neuron-astrocyte network models.

(A) Strong scaling. In this type of benchmark, the network size stays constant (20,000 cells), but the number of compute nodes used is scaled. (B) Weak scaling. In this type of benchmark, the model size and the number of compute nodes are scaled together. The first row in (A) and (B) shows network connection time and state propagation time for a model time of $T_{\text{model}}=10 \text{s}$ and its real-time factor ($T_{\text{wall}}$/$T_{\text{model}}$). Error bars indicate standard deviations across nine simulations. In some cases the error bar is very small and hence not visible. The second row shows the phases of state propagation in terms of average real-time factor. “Update”: advancing the dynamical states in each cell and device; “Spike CCD”: collocation, communication, and delivery of spikes; “SIC GD”: gathering and delivery of SIC events; “Other”: the part of time that cannot be attributed to the three phases above. “Sparse”, “Synchronous”: models with sparse and synchronous activities, respectively. “Surrogate”: Model using astrocyte_surrogate (no astrocytic dynamics). For specifications of these models, see Tables A and B in S1 Appendix. For the dynamics of the “Sparse” and “Synchronous” models, see Fig 4.