Performance and Energy Efficiency on Llama 8B (32K Context)
1TTFT (s)SIFT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SIFTMode=SIFT2026.06 | 1 | 7.87 | 0.127 | 7.83 | 7.8 | |
| RecomputeMode=Recompute2026.06 | 1.67 | 8.82 | 0.113 | 14.7 | 24.47 | |
| CacheBlendMode=CacheBlend2026.06 | 2.13 | 5.51 | 0.181 | 11.72 | 24.9 |