Image Encoder Inference on CLIP
2.15Latency (ms)ELSA-Turbo (Ours)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ELSA-Turbo (Ours)Model=ViT-B/32, Precision=TF32-Turbo, Scope=image_encoder only2026.04 | 2.15 | 1.115 | |
| ELSA-Strict (Ours)Model=ViT-B/32, Precision=strict FP32, Scope=image_encoder only2026.04 | 2.159 | 1.111 | |
| ME-SDPA (baseline)Model=ViT-B/32, Scope=image_encoder only2026.04 | 2.398 | 1 | |
| ELSA-Turbo (Ours)Model=ViT-B/16, Precision=TF32-Turbo, Scope=image_encoder only2026.04 | 4.329 | 1.055 | |
| ELSA-Strict (Ours)Model=ViT-B/16, Precision=strict FP32, Scope=image_encoder only2026.04 | 4.365 | 1.046 | |
| ME-SDPA (baseline)Model=ViT-B/16, Scope=image_encoder only2026.04 | 4.568 | 1 | |
| ELSA-Turbo (Ours)Model=ViT-L/14, Precision=TF32-Turbo, Scope=image_encoder only2026.04 | 14.842 | 1.039 | |
| ELSA-Strict (Ours)Model=ViT-L/14, Precision=strict FP32, Scope=image_encoder only2026.04 | 14.914 | 1.034 | |
| ME-SDPA (baseline)Model=ViT-L/14, Scope=image_encoder only2026.04 | 15.419 | 1 | |
| ELSA-Turbo (Ours)Model=ViT-L/14@336px, Precision=TF32-Turbo, Scope=image_encoder only2026.04 | 27.931 | 1.062 | |
| ELSA-Strict (Ours)Model=ViT-L/14@336px, Precision=strict FP32, Scope=image_encoder only2026.04 | 27.989 | 1.06 | |
| ME-SDPA (baseline)Model=ViT-L/14@336px, Scope=image_encoder only2026.04 | 29.665 | 1 |