From: Neural network training with limited precision and asymmetric exponent
Bit counts | Network accuracy | ||||
---|---|---|---|---|---|
Bit count | Exponent | Mantissa | Limitation only (%) | Limitation with asymmetric exponent (%) | Limitation with asymmetric exponent (excluding gradients) (%) |
5 | 3 | 1 | 12.08 | 10.09 | 9.80 |
6 | 3 | 2 | 9.39 | 8.92 | 14.29 |
7 | 3 | 3 | 10.42 | 9.58 | 57.31 |
8 | 3 | 4 | 6.63 | 10.10 | 75.89 |
9 | 3 | 5 | 7.17 | 22.42 | 85.97 |
10 | 3 | 6 | 16.37 | 41.32 | 89.31 |
11 | 3 | 7 | 54.61 | 66.52 | 90.27 |
12 | 3 | 8 | 71.45 | 86.98 | 95.01 |
13 | 3 | 9 | 85.77 | 91.34 | 94.07 |
14 | 3 | 10 | 91.06 | 94.54 | 97.13 |
15 | 3 | 11 | 93.26 | 95.93 | 97.76 |
16 | 3 | 12 | 94.77 | 97.29 | 97.32 |