Fig. 4
Neuroscience use case. (A) Left: Sketch of a model40 describing the activity dynamics generated by a local neuronal circuit of the mammalian neocortex (adapted from42). The network model is composed of four excitatory (E; blue triangles) and four inhibitory neuronal populations (I; red circles); distributed across four cortical layers 2/3, 4, 5 and 6; and driven by background inputs. Neurons in the network are interconnected in a cell-type and layer specific manner (blue and red arrows). Right: The model generates neuronal activity data (“spikes” and firing rates) as the primary neuroscientific data (blue cylinder). For each simulation instance, information about the model parameterization, random number generator (RNG) seeds, the hard- and software configuration, as well as the wall-clock times are stored in various files and formats as raw metadata (yellow cylinder). In a subsequent post-processing step (red gear), the metadata is parsed and structured by the Archivist. The simulation data is annotated with this structured metadata and stored in a database for further usage (red cylinder). The database can flexibly be queried according to user interests (curved red arrows). (B, C) Verification (B) and performance benchmarking (C) as two exemplary types of data usage. (B) Average activity level (firing rate) in each of the 8 neuronal populations 2/3E, …, 6I depicted in panel A, obtained from simulations of the model on four different GPU platforms (see labels at horizontal axis in panel (C). (C) Real time factor (ratio between wall-clock time Twall and simulated biological time Tmodel = 10 s) for four different GPU computing platforms. Error bars (red) in (B) and (C) depict standard deviations across ten different model realizations (random-number generator [RNG] seeds) and simulation runs for each platform (error bars are partly too small to be visible).
