Table 4.
Results comparison for MNIST handwritten digit classification task in various hardware architectures.
| System | Dataset | Accuracy | Inference | Inference | Silicon |
|---|---|---|---|---|---|
| (%) | Energy (μJ) | Time (ms) | Area (mm2) | ||
| Event-Driven CNN | MNIST | 99.44 | 12 | 1.1 | 1.88 |
| (SENECA) | |||||
| Spiking CNN (LOIHI) | MNIST | 99.21 | 660 | 6.65 | 5.74 |
| (Rueckauer et al., 2022) | N-MNIST | 98.51 | 620 | 7.07 | 12.3 |
| Spiking CNN (Speck) | N-MNIST | 98.56 | 180 | 300 | 30 |
| (Richter et al., 2023) | |||||
| Spiking CNN (TrueNorth) | MNIST | 99.42 | 108 | 4 | 192 |
| (Esser et al., 2015) | |||||
| CNN (MAX78000) | MNIST | 99.44 | 215 | 8.3 | NA |
| (Moss et al., 2022) |
The Silicon area is the total area of the utilized cores.