Table 3 Comparison of the energy consumption per layer of the models for one forward pass as presented in Fig. 6

From: Spiking neural networks for nonlinear regression of complex transient signals on sustainable neuromorphic processors

Layers

GPU (nJ)

CPU (nJ)

Loihi (nJ)

Loihi + GPU (nJ)

Loihi + CPU (nJ)

LMU/SLMU layer 1

3434

99081

5.4072

5.4072

5.4072

LMU/SLMU layer 2

6331

181641

5.382

5.382

5.382

LMU/SLMU layer 3

6331

181641

5.385

5.385

5.385

RNN/SRNN layer

3100

90000

0.075

0.075

0.075

Dense layer 1

610

18000

0.027

610

18000

Dense layer 2

310

8800

0.011

310

8800

Dense layer 3

9.6

280

0.00032

9.6

280

Total Energy

20125.6

579443

16.285

945.862

27096.262

  1. The Spiking LMU (SLMU) consumes only a small percentage compared to LMU when deployed on a neuromorphic chip, e.g. 5.4072nJ for one forward pass through layer 1 with Loihi compared to 99081nJ on a CPU.
Back to article page