Table 2.
Communication and computation delays in each layer.
Layer (Input Size) | Operator | Stride | PWC+DWC | PWC | ||
---|---|---|---|---|---|---|
Communication Delay (us) | Computation Delay (us) | Communication Delay (us) | Computation Delay (us) | |||
224 × 224 × 3 | Conv2d | 2 | N/A | N/A | 24.56 | 405.81 |
112 × 112 × 32 | Bottleneck | 1 | 31.40 | 586.67 | 24.56 | 404.49 |
112 × 112 × 16 | Bottleneck | 2 | 34.01 | 1296.53 | 65.02 | 270.47 |
1 | 30.65 | 600.87 | 108.00 | 430.33 | ||
56 × 56 × 24 | Bottleneck | 2 | 30.65 | 522.08 | 106.08 | 129.47 |
1 | 24.07 | 265.79 | 127.01 | 150.92 | ||
28 × 28 × 32 | Bottleneck | 2 | 24.07 | 171.47 | 33.04 | 78.68 |
1 | 16.55 | 209.50 | 65.68 | 144.35 | ||
14 × 14 × 64 | Bottleneck | 1 | 16.55 | 209.47 | 65.79 | 216.36 |
1 | 24.71 | 412.38 | 98.43 | 314.83 | ||
14 × 14 × 96 | Bottleneck | 2 | 24.71 | 341.67 | 28.08 | 152.12 |
1 | 21.43 | 309.20 | 46.32 | 244.99 | ||
7 × 7 × 160 | Bottleneck | 1 | 21.43 | 309.17 | 46.85 | 489.72 |
7 × 7 × 320 | Conv2d | 1 | N/A | N/A | 19.65 | 670.02 |
1280 × 1000 | Conv2d | 1 | N/A | N/A | 9.77 | 1345.96 |
1 Include pooling or concatenation if necessary.