Attention Kernel Performance Evaluation on Sparse-Attention Micro-Benchmark N=600 (test)
0.42Latency (ms)TempoNet
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TempoNetPattern=deadline-aware block + chunk top-k, Scheduling-specific?=Yes2026.02 | 0.42 | 0.62 | |
| Efficient Sparse AttentionPattern=learned block-drop, Scheduling-specific?=No2026.02 | 0.65 | 0.95 | |
| Explicit Sparse TransformerPattern=static column-drop, Scheduling-specific?=No2026.02 | 0.68 | 1 |