Time-to-accuracy Language Modeling on GPT-2 Medium (train)
29.47Training Time (h)PACI
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PACIBatch=256, PPL threshold=≤ 18, Num. micro=162026.06 | 29.47 | 1.24 | 1.24 | |
| FlushBatch=256, PPL threshold=≤ 18, Num. micro=162026.06 | 36.57 | 1 | 1 | |
| PACIBatch=128, PPL threshold=≤ 18, Num. micro=82026.06 | 37.32 | 1.69 | 1.55 | |
| PACIBatch=256, PPL threshold=≤ 18, Num. micro=82026.06 | 40.16 | 1.24 | 0.91 | |
| PACIBatch=256, PPL threshold=≤ 17, Num. micro=162026.06 | 45.15 | 1.33 | 1.33 | |
| FlushBatch=256, PPL threshold=≤ 18, Num. micro=82026.06 | 49.85 | 1 | 0.73 | |
| PACIBatch=128, PPL threshold=≤ 18, Num. micro=42026.06 | 50.57 | 1.75 | 1.14 | |
| PACIBatch=128, PPL threshold=≤ 17, Num. micro=82026.06 | 52.86 | 1.74 | 1.59 | |
| FlushBatch=128, PPL threshold=≤ 18, Num. micro=162026.06 | 57.72 | 1 | 1 | |
| FlushBatch=256, PPL threshold=≤ 17, Num. micro=162026.06 | 59.99 | 1 | 1 | |
| PACIBatch=256, PPL threshold=≤ 17, Num. micro=82026.06 | 60.18 | 1.36 | 1 | |
| FlushBatch=128, PPL threshold=≤ 18, Num. micro=82026.06 | 63.03 | 1 | 0.92 | |
| PACIBatch=256, PPL threshold=≤ 16, Num. micro=162026.06 | 64 | 1.36 | 1.36 | |
| PACIBatch=128, PPL threshold=≤ 17, Num. micro=42026.06 | 69.89 | 1.85 | 1.21 | |
| PACIBatch=128, PPL threshold=≤ 16, Num. micro=82026.06 | 70.31 | 1.8 | 1.64 | |
| FlushBatch=256, PPL threshold=≤ 17, Num. micro=82026.06 | 81.77 | 1 | 0.73 | |
| PACIBatch=256, PPL threshold=≤ 16, Num. micro=82026.06 | 84.11 | 1.42 | 1.04 | |
| FlushBatch=128, PPL threshold=≤ 17, Num. micro=162026.06 | 84.29 | 1 | 1 | |
| FlushBatch=256, PPL threshold=≤ 16, Num. micro=162026.06 | 87.32 | 1 | 1 | |
| FlushBatch=128, PPL threshold=≤ 18, Num. micro=42026.06 | 88.39 | 1 | 0.65 | |
| FlushBatch=128, PPL threshold=≤ 17, Num. micro=82026.06 | 92.05 | 1 | 0.92 | |
| PACIBatch=128, PPL threshold=≤ 16, Num. micro=42026.06 | 92.92 | 1.91 | 1.24 | |
| FlushBatch=128, PPL threshold=≤ 16, Num. micro=162026.06 | 115.63 | 1 | 1 | |
| FlushBatch=256, PPL threshold=≤ 16, Num. micro=82026.06 | 119.02 | 1 | 0.73 | |
| FlushBatch=128, PPL threshold=≤ 16, Num. micro=82026.06 | 126.28 | 1 | 0.92 | |
| FlushBatch=128, PPL threshold=≤ 17, Num. micro=42026.06 | 129.08 | 1 | 0.65 | |
| FlushBatch=128, PPL threshold=≤ 16, Num. micro=42026.06 | 177.07 | 1 | 0.65 |