Fig. 5: Power and efficiency analysis.
From: Demonstration of transformer-based ALBERT model on a 14nm analog AI inference chip
a Simulated time sequence of activities in the four layer-blocks implemented in the analog accelerator (inProj, outProj, FC1, and FC2), attention computation ("QKV''), and all other digital operations. b–e show the dependence on the mean sequence-length of: (b) the ratio between analog and digital operation numbers, (c) the ratio between analog and digital energy, (d) throughput in samples-per-second, and (e) energy efficiency in samples-per-Joule. f The percentage of total energy spent on analog operations, (g) throughput, (h) task-based energy efficiency, and (i) system-level energy-efficiency in TOPS/W, across the 7 GLUE tasks.
