Table 23 Computational performance comparison across sample sizes and architectures.
Architecture | Sample size (K) | Training time (hours) | Memory usage (GB) | GPU memory (GB) | Convergence epochs | Time per epoch (min) | Total GPU hours | Energy consumption (kWh) |
|---|---|---|---|---|---|---|---|---|
Dual-Gland GAN (Ours) | 10 | 8.4 | 12.3 | 18.7 | 78 | 6.5 | 8.4 | 21.8 |
25 | 18.7 | 28.4 | 24.1 | 82 | 13.7 | 18.7 | 48.6 | |
50 | 34.2 | 45.6 | 31.2 | 85 | 24.1 | 34.2 | 89.1 | |
100 | 62.8 | 78.9 | 42.3 | 89 | 42.3 | 62.8 | 163.4 | |
Traditional GAN | 10 | 14.6 | 8.9 | 14.2 | 156 | 5.6 | 14.6 | 38.0 |
25 | 35.8 | 18.7 | 19.3 | 168 | 12.8 | 35.8 | 93.2 | |
50 | 78.4 | 32.1 | 25.6 | 184 | 25.6 | 78.4 | 204.1 | |
100 | 167.2 | 58.3 | 34.7 | 201 | 49.9 | 167.2 | 435.1 | |
DCGAN | 10 | 11.2 | 9.8 | 15.6 | 125 | 5.4 | 11.2 | 29.1 |
25 | 26.4 | 21.3 | 21.8 | 134 | 11.8 | 26.4 | 68.6 | |
50 | 58.7 | 36.9 | 28.4 | 147 | 24.0 | 58.7 | 152.6 | |
100 | 128.5 | 67.2 | 38.1 | 159 | 48.5 | 128.5 | 334.1 | |
WGAN | 10 | 9.8 | 10.4 | 16.2 | 108 | 5.4 | 9.8 | 25.5 |
25 | 22.1 | 23.6 | 22.9 | 115 | 11.5 | 22.1 | 57.5 | |
50 | 47.3 | 41.2 | 29.7 | 124 | 22.9 | 47.3 | 123.0 | |
100 | 98.7 | 75.8 | 39.4 | 132 | 44.9 | 98.7 | 256.6 | |
ECP-IGANN | 10 | 12.8 | 14.7 | 22.3 | 95 | 8.1 | 12.8 | 33.3 |
25 | 31.5 | 32.1 | 28.6 | 102 | 18.5 | 31.5 | 82.0 | |
50 | 68.9 | 58.4 | 36.2 | 109 | 37.9 | 68.9 | 179.1 | |
100 | 145.7 | 103.6 | 47.8 | 118 | 74.1 | 145.7 | 378.8 | |
GSIP-GAN | 10 | 15.3 | 16.2 | 24.1 | 112 | 8.2 | 15.3 | 39.8 |
25 | 37.8 | 35.7 | 31.4 | 118 | 19.2 | 37.8 | 98.3 | |
50 | 82.4 | 64.3 | 39.7 | 125 | 39.6 | 82.4 | 214.2 | |
100 | 174.6 | 118.9 | 52.1 | 133 | 78.7 | 174.6 | 454.0 |
