Attention Operator Performance on MLA Sequence Length 512 128 Head Dimension
50.6TFLOPSLLM-TL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LLM-TLHead Dimension=128, Mask=Causal, GPU=A100, Base Model=DeepSeek-V32025.06 | 50.6 | 2.86 | |
| cuDNNHead Dimension=128, Mask=Causal, GPU=A1002025.06 | 35.5 | — | |
| torchHead Dimension=128, Mask=Causal, GPU=A1002025.06 | 22.9 | — | |
| DeepSeek-V3Head Dimension=128, Mask=Causal, GPU=A1002025.06 | 17.7 | — |