Model Efficiency and Performance on Llama2 7B (Language)
2.05Peak PerformanceISSCC 25
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ISSCC 25Technology=28nm, Die Area=12.66mm2, Frequency=10-230MHz, Supply Voltage=0.57-0.9V, Architecture Type=Heterogeneous, Outlier Handling Methods=Outlier-Aware Mixed Precision, Quant. Precision=IA: FP12/INT9, W: INT4/8, Quant. Granularity=Per-Vector2025.12 | 2.05 | — | — | 13.4 | — | |
| SeVeDoTechnology=28nm, Die Area=10.72mm2, Frequency=250MHz, Supply Voltage=0.9V, Architecture Type=Heterogeneous, Outlier Handling Methods=SVD-based Decomp. + Group Quantization, Quant. Precision=(Res) IA: INT4, W: INT4 / (LR) IA: INT16/8, W: INT8/4, Quant. Granularity=(Res) Per-G128/SubG32, (LR) Per-Vector2025.12 | — | 8.19 | 0.612 | — | 13.4 |