Extended Data Fig. 3: Diffusion Random Walks Scaling Studies.

(A) Walker updates per second for a 1,000 (dark green) and 32,000 (light green) basic diffusion simulation across conventional and neuromorphic platforms. (B) Comparison of Loihi and single-chip TrueNorth to a single-core CPU simulation on normalized time of a simple diffusion simulation as a function of increasing random walkers. All times normalized to the time it takes to complete a simulation with 1,000 walkers. (C) Comparison of multi-chip TrueNorth to multi-core CPU and GPU simulations. GPU generates threads for all walker scenarios; GPU Single Block allocates only 1,024 threads for all walkers. (D) TrueNorth Execution time reaches a limit as mesh counts increase. (E) TrueNorth Execution time scales linearly with walker count, again, but also demonstrates the sensitivity of the algorithm to bottlenecks caused by uneven transition probabilities. (F) TrueNorth Execution time is dramatically reduced once all walkers do not start on the same position. (G) Time required for NVIDIA Titan XP GPU to simulate diffusion on a torus for 100,000 time steps as a function of the number of walkers. A fixed 1,024 threads are allocated for each trial. (H) Time required for NVIDIA Titan XP GPU to simulate diffusion on a torus for 100,000 time steps as a function of the number of walkers. For this weak-scaling experiment, a block of 1024 is added for every 1,000 walkers. (I) Time required for NVIDIA Titan XP GPU to simulate diffusion on a torus for a single time step as a function of the number of walkers. For this weak-scaling experiment, a block of 1024 is added for every 1,000 walkers.