Table 1 Performance comparison of our approach to conventional computing systems and other optical/optoelectronic approaches
From: Low-power scalable multilayer optoelectronic neural networks enabled with incoherent light
Technique | Approach | Throughput | Efficiency (Expt) | Efficiency (Proj) | Precision | Reference |
|---|---|---|---|---|---|---|
TOPS | TOPS/W | TOPS/W | bit | |||
NVIDIA B200 | GPU | 144 ⋅ 103* | 10.01 | 4 | ||
57 ⋅ 103* | 5.03 | 8 | ||||
NVIDIA RTX 4090 | GPU | 660.60** | 0.78 | 8 | ||
Google TPUv4 | ASIC | 275 | 1.62 | 8 | ||
Photonic WDM/PCM in-memory computing | Photonic | 0.65 | 0.50 | 7.00 | 5 | |
Image intensifier | Incoherent Free Space | 5.76 ⋅ 10−7 | 3.03 ⋅ 10−7 | 66.67 | 8 | |
Photonic convolutional accelerator | Photonic | 0.48 | 1.26 | 8 | ||
Free space optoelectronic neural network | Incoherent Free Space | 1.6 ⋅ 10−3 | 11.45 ⋅ 10−3 | 35.09 | 8 | This Work |
