Transformer Inference on Spartan-7 XC7S15 FPGA
3.77RMSEResource-aware Mixed-precision Quantization
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Resource-aware Mixed-precision Quantizationn (Sequence length)=12, Combinations (Mixed-precision bit-widths)=8, 8, 6, 8, 6, 4, 8, 8, 8, 82024.10 | 3.77 | 5.78 | 31 | 39 | 70 | 0.405 | |
| Resource-aware Mixed-precision Quantizationn (Sequence length)=24, Combinations (Mixed-precision bit-widths)=6, 8, 4, 4, 4, 4, 4, 4, 8, 82024.10 | 4.1 | 11.92 | 31 | 37 | 68 | 0.811 | |
| Resource-aware Mixed-precision Quantizationn (Sequence length)=18, Combinations (Mixed-precision bit-widths)=8, 6, 4, 4, 6, 4, 4, 4, 8, 82024.10 | 4.19 | 8.79 | 31 | 37 | 68 | 0.598 |