Extended Data Table 4 Details of the YOLO neural networks used in the demonstration
From: A dual-domain compute-in-memory system for general neural network inference
Description | YOLO w/o enhancing | YOLO w/ enhancing | YOLO w/ enhancing (group conv.) |
|---|---|---|---|
Feature extraction layers (ACIM) | 11 | 11 | 11 |
Detection head layers (digital) | 2 | 2 | 2 |
Feature-enhancing layers (digital) | 0 | 11 | 11 |
Number of layers | 13 | 24 | 24 |
Layers on ACIM | 11 | 11 | 11 |
Number of weight parameters (M) | 0.75 | 0.87 | 0.76 |
Number of weight parameters on ACIM (M) | 0.74 | 0.74 | 0.74 |
Total GOPs | 0.83 | 0.96 | 0.84 |
Enhancing layer GOPs | 0 | 0.13 | 0.012 |
Enhancing layer extra overhead (%) | 0 | 13 | 1.4 |
Est. 28 nm energy efficiency (TOPs W−1) | 14.1 | 12.4 | 13.7 |
Est. 28 nm energy efficiency (TOPs W−1) (consider enhancing layers as invalid operations) | – | 10.5 | 12.9 |
mAP-50 (on-chip measured) | 0.27 | 0.724 | – |
mAP-50 (simulated) | 0.23 | 0.744 | 0.662 |
mAP-50 Improvement with Enhancing | - | ×2.7 | ×2.5 |
