Inference Efficiency Analysis on MSCOCO Karpathy
9.7End-to-End Latency (ms)EfficientVLM
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| EfficientVLMInput Length (image/text)=196/35, Hardware=GPU2022.10 | 9.7 | 42 | 5 | 8.3 | 50.3 | 8.5 | 1.3 | |
| X-VLMclipInput Length (image/text)=196/35, Hardware=GPU2022.10 | 17.8 | 86.1 | 9 | 18.9 | 123 | — | 4.2 | |
| ViLTInput Length (image/text)=200/40, Hardware=GPU2022.10 | 21 | 2.4 | 0.7 | 0.6 | 109 | 20.3 | 22.8 | |
| MiniVLMInput Length (image/text)=50/35, Hardware=GPU2022.10 | 23.6 | 7.5 | 12.2 | 4.4 | 34.5 | 11.4 | 2.3 | |
| DistilVLMInput Length (image/text)=50/35, Hardware=GPU2022.10 | 23.6 | 7.5 | 12.2 | 4.4 | 34.5 | 11.4 | 2.3 | |
| ViLTInput Length (image/text)=200/40, Hardware=CPU2022.10 | 69.1 | 2.4 | 0.8 | 0.6 | 109 | 68.3 | 22.8 | |
| OSCAR_BInput Length (image/text)=50/35, Hardware=GPU2022.10 | 135.2 | 63.8 | 121.9 | 767 | 109 | 13.3 | 8.2 | |
| EfficientVLMInput Length (image/text)=196/35, Hardware=CPU2022.10 | 180.8 | 42 | 171.5 | 8.3 | 50.3 | 17.1 | 1.3 | |
| X-VLMclipInput Length (image/text)=196/35, Hardware=CPU2022.10 | 395.5 | 86.1 | 355.1 | 18.9 | 123 | — | 4.2 | |
| MiniVLMInput Length (image/text)=50/35, Hardware=CPU2022.10 | 418.2 | 7.5 | 393.9 | 4.4 | 34.5 | 24.3 | 2.3 | |
| DistilVLMInput Length (image/text)=50/35, Hardware=CPU2022.10 | 418.2 | 7.5 | 393.9 | 4.4 | 34.5 | 24.3 | 2.3 | |
| OSCAR_BInput Length (image/text)=50/35, Hardware=CPU2022.10 | 12,347.1 | 63.8 | 12,300 | 767 | 109 | 47.1 | 8.2 |