Table 2.
Array sizes, weight sharing factors and number of MACs performed for each layer for AlexNet* (Krizhevsky et al., 2012) architecture.
| Layer | RPU array size (matrix size) | Weight sharing factor (ws) | MACs |
|---|---|---|---|
| K1 | 96 × 363 | 3, 025 | 106 M |
| K2 | 256 × 2, 400 | 729 | 448 M |
| K3 | 384 × 2, 304 | 169 | 150 M |
| K4 | 384 × 3, 456 | 169 | 224 M |
| K5 | 256 × 3, 456 | 169 | 150 M |
| W6 | 4, 096 × 9, 216 | 1 | 38 M |
| W7 | 4, 096 × 4, 096 | 1 | 17 M |
| W8 | 1, 000 × 4, 096 | 1 | 4 M |
| Total MACs = 1.14 G |
*
Table assumes the weights that are originally distributed to two GPUs are contained into a single RPU array for each layer.