Table 3.
Comparison of simulations with different hardware constraints.
| No | Description | Rafter | DKS | Reference |
|---|---|---|---|---|
| 1 | Baseline* | 0.54 ± 0.05 | – | – |
| 2 | Baseline with noise† | 0.45 ± 0.03 | – | – |
| DISCRETIZED WEIGHTS | ||||
| 3 | 8 bit | 0.53 ± 0.03 | 0.008* | (1) |
| 4 | 6 bit | 0.52 ± 0.03 | 0.039* | (1) |
| 5 | 4 bit, deterministic | 0.37 ± 0.03 | 0.098* | (1) |
| 6 | 4 bit, probabilistic | 0.46 ± 0.03 | 0.053† | (2) |
| THRESHOLD READOUT | ||||
| 7 | 8 bit | 0.59 ± 0.03 | 0.140* | (1) |
| 8 | 6 bit | 0.59 ± 0.05 | 0.120* | (1) |
| 9 | 4 bit, deterministic | 0.27 ± 0.04 | 0.154* | (1) |
| 10 | 4 bit, probabilistic | 0.48 ± 0.05 | 0.043† | (2) |
The table lists the final performance Rafter and the Kolmogorov–Smirnov (KS) measure DKS comparing the final weight distribution to the one of the reference simulation indicated by its row number in the last column. For this comparison, the continuous reference distribution is rounded to the respective resolution.
The p-value for the KS test is p = 0.35 for 8 bit and p < 0.01 in all other cases.