Fig. 3: Comparison of scalability in NeuroScale, Loihi, and TrueNorth.

a Wall-clock times for systems with increasing number of cores. As the number of cores increases, the wall-clock time of NeuroScale remains steady (\({{{\mathcal{O}}}}(1)\) scaling), while the wall-clock times of Loihi and TrueNorth gradually increase (\({{{\mathcal{O}}}}(\sqrt{n})\) scaling, where n is the core count). Wall-clock times are normalized to the wall-clock time of a 256-core NeuroScale system. For a 16,384-core system, NeuroScale is over four times faster than Loihi and TrueNorth. Note that Intel’s recently announced Hala Point system has 140,544 cores. b Wall-clock times of a fixed-size network spanning 256 cores on systems of increasing size. As the system size increases, the wall-clock time of NeuroScale stays constant (\({{{\mathcal{O}}}}(1)\) scaling), while that of Loihi and TrueNorth keeps increasing (\({{{\mathcal{O}}}}(\sqrt{n})\) scaling). c Effect of locality on the speedup of NeuroScale compared to Loihi and TrueNorth. A more significant speedup is achieved when communicating cores are placed in closer physical proximity (higher locality). d Effect of activity sparsity. NeuroScale achieves higher speedup with sparser spiking activity. Sparsity is measured as the average time interval between spikes.