Table 5. Accuracy of the analog inference accelerator on the full ImageNet test set (50,000 images, 10 runs each). ResNet50-v1.5 Fully digital Design A, ideal cells Design A, SONOS Floating-point 76.466% 76.082% 74.296% ±0.348% 4-bit, QAT 76.154% 76.038% 75.294% ±0.192% is state-proportional error, as explained in Section 9.1. Table 5 further shows that by using a standard quantization-aware training scheme with 4-bit activations, the propagation of this error can be significantly suppressed as described in Section 7. This reduces the accuracy loss induced by the SONOS device to only 0.86%. 10. Conclusions Error resilience can be built into analog accelerators by