Table 1 Performance comparison of our proposed OCPU framework
From: Compact optical convolution processing unit based on multimode interference
Type | Programmable units | Matrix dimension | Platform | Accuracy on MNIST test set | Network architecture | Efficiency (/MAC) | Precision of results | Compute density (MACs/s/mm2) | Scale |
|---|---|---|---|---|---|---|---|---|---|
TOPS-CA35 | / | 9 × 10 | / | 88.00% | 1Conv. (3 5 × 5 kernels) + 1FC | 1.58 pJ | 7-bit | System | / |
Netcast52 | / | / | / | 98.80% | 3FC | 10.00 fJe | 8-bit | System | / |
AOM-VMM53 | 3 | 1 × 3 | / | 98.90% | 2Conv. (16 3 × 3 kernels) + 2FC | / | / | System | N2 |
MZI-VMM25 | 60 | 4 × 4 | Si | 76.70% (4 categories, vowel recognition) | 2FC | 30.00 fJa | 5-bit | 0.56 Ta | N2 |
MRR-VMM33 | 16 | 4 × 4 | Si | / | / | 0.18 pJb | 4-bitf | 1.60 Tb | N2 |
MRR-VMM51 | 4 | 1 × 4 | Si | 97.41% | 3FC | 0.56 pJb | 4-bit | 2.89 Tb | N2 |
PCM-VMM16 | 36 | 9 × 4 | SiN | 95.30% | 1Conv. (4 2 × 2 kernels) + 1FC | 5.00 pJ | 7-bitd | 0.60 T | N2 |
64 | 8 × 8 | Si | / | / | 4.00 pJ | 7-bitd | 81.00 T | N2 | |
PMMC-VMM54 | 4 | 1 × 4 | SiN | 91.00% (2 categories) | 1Conv. (2 2 × 2 kernels) + 1FC | / | 6-bitf | 82.00 Tb | N2 |
IDNN-VMM24 | 20 | 10 × 10c | Si | 89.40% | 2FC | / | / | / | 2N |
Flash (analog electronics, simulation)55 | / | 100 × 100 | Si | / | / | 7.00 fJ | 5-bit | 18.00 T | / |
This work | 4 | 4 × 4c | SiN | 92.17% | 1Conv. (2 2 × 2 kernels) + 1FC | 4.84 pJ | 5-bit | 12.74 T | N |
Expected from this work | 9 | 9 × 9c | Si | 96.35% | 1Conv. (8 3 × 3 kernels) + 1FC | 0.95 pJ | 5-bit | 1.19 P | N |
