Image Classification on ImageNet-1K (Top-1 Accuracy & OOD Robustness)
85Top-1 AccuracyH2(ViT-H/14)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| H2(ViT-H/14)params=474.2, throughput=640, bf16 FLOPs=127.4, Peak Mem.=27342025.05 | 85 | -13.4 | |
| H8(ViT-H/14)params=355.8, throughput=660, bf16 FLOPs=102.3, Peak Mem.=22232025.05 | 85 | -13.4 | |
| I8(ViT-H/14)params=362.3, throughput=657, bf16 FLOPs=104.0, Peak Mem.=22492025.05 | 84.7 | 0 | |
| ViT-H/14params=632.1, throughput=569, bf16 FLOPs=167.8, Peak Mem.=32852025.05 | 84.6 | -12.6 | |
| H2(ViT-L/16)params=228.3, throughput=1610, bf16 FLOPs=46.9, Peak Mem.=14582025.05 | 84.5 | -13.9 | |
| H8(ViT-L/16)params=171.3, throughput=1615, bf16 FLOPs=37.7, Peak Mem.=11942025.05 | 84.5 | -13.6 | |
| ViT-L/16params=304.4, throughput=1421, bf16 FLOPs=61.9, Peak Mem.=15572025.05 | 84.2 | -13.4 | |
| I8(ViT-L/16)params=175.5, throughput=1618, bf16 FLOPs=38.5, Peak Mem.=12102025.05 | 84 | 0 |