Fig. 4: Hardware (HW) demonstration of HWA training, drift effects, and MAC variation.
From: Demonstration of transformer-based ALBERT model on a 14nm analog AI inference chip
a Experimentally measured HW inference accuracy with optimized (grey bars) and without any (white bars) HWA noise during fine-tuning of seven GLUE tasks. Adding the appropriate level of noise during fine-tuning improves the average HW accuracy of seven GLUE tasks by 4.4%. b The dependence of HW accuracy on the noise scale used in HWA fine-tuning for six GLUE tasks (MRPC: blue squares; CoLA: lime stars; SST-2: pink triangles; QNLI: aqua triangles; QQP: black diamonds; MNLI: brown circles). The smallest task (RTE) is excluded due to its instability. The filled-in (enlarged) symbol identifies the optimal HWA noise-scale for that task, typically in the range of 1.0–2.0. c The HW inference accuracy of the MRPC task with (red squares) and without (blue circles) recalibration over 30 days. d The distribution of MAC error from day 0 up to 3 weeks after weight programming. e The ratio of the median MAC variation over median MAC values for repeated inference attempts.
