Fig. 5: PACE Ising convergence and its comparison with a commercial A10 GPU.
From: An integrated large-scale photonic accelerator with ultralow latency
a, PACE Ising convergence rates versus latency. The error bars denote the standard deviation of ten random tests. Inset, non-convergence when the light is off in the circuit, indicating the effectiveness of oMAC and that pure hardware noise alone cannot induce convergence. b, Statistics of the number of iterations to solution at 5 ns latency, over 2,000 random input states. Inset, minima energy evolution in the iterative search. Dashed red line denotes the target minima energy. c, Statistics of the number of iterations that converge at 2 ns latency in the PACE system. Inset, energy evolution over 5,000 iterations for an unconverged case. Dash red line denotes the target minima energy. d, Comparison of the number of iterations to solution between the PACE system and the commercial GPU. The average counts are 537 for the PACE system and 347 for the NVIDIA A10 GPU, shown by dashed lines. Inset, comparison of single cycle latency and total time consumption between the PACE system and the NVIDIA A10 GPU. PACE data are multiplied by 100 for visual comparison.
