Image Classification on ImageNet 1K (Throughput and Memory Usage)
1,309.24Throughput (img/s)ELSA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ELSAModel=ViT-T, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 1,309.24 | 33.8 | 22.3 | |
| ELSAModel=ViT-S, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 1,275.71 | 74 | 23.2 | |
| ELSAModel=ViT-M, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 1,204.31 | 113.3 | 25.6 | |
| ELSAModel=ViT-B, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 1,064.34 | 220.8 | 21.8 | |
| Math-SDPAModel=ViT-S, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 897.03 | 96.4 | — | |
| Math-SDPAModel=ViT-T, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 868.44 | 43.5 | — | |
| FA2Model=ViT-S, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 859.73 | 79.8 | 17.2 | |
| Math-SDPAModel=ViT-M, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 838.34 | 152.2 | — | |
| FA2Model=ViT-M, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 828.39 | 130.3 | 14.4 | |
| Math-SDPAModel=ViT-B, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 820.91 | 282.3 | — | |
| FA2Model=ViT-T, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 791.03 | 36 | 17.2 | |
| FA2Model=ViT-B, Attention Type=Global, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 787.39 | 249.6 | 11.6 | |
| W-ELSAModel=Swin-T/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 597.23 | 232.1 | 14.8 | |
| FA2Model=Swin-T/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 526.65 | 223.2 | 10.4 | |
| W-ELSAModel=Swin-T/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 520.22 | 351.9 | 3.2 | |
| WAModel=Swin-T/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 493.2 | 341 | — | |
| FA2Model=Swin-T/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 485.45 | 351.9 | 3.2 | |
| WAModel=Swin-T/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 479.98 | 202.2 | — | |
| W-ELSAModel=Swin-S/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 305.22 | 369.1 | 5.7 | |
| FA2Model=Swin-S/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 267.89 | 401.8 | 2.7 | |
| W-ELSAModel=Swin-S/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 266.59 | 291.3 | 18.6 | |
| FA2Model=Swin-S/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 260.23 | 267.5 | 8.9 | |
| WAModel=Swin-S/W8, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 241.9 | 245.7 | — | |
| WAModel=Swin-S/W16, Attention Type=Window, Precision=FP16, Batch size=8, Hardware=A100-40GB, Resolution=224x224, Scan Protocol=strict scan only2026.04 | 241.77 | 391.3 | — |