Table 1 Energy consumed in mJ by the whole STPN network during one game of Atari Pong

From: Single neuromorphic memristor closely emulates multiple synaptic mechanisms for energy efficient neural networks

  

ΔF

Decay

W + F

ST-Hebb

weight mult.

Total [mJ]

Memristor

 

0.4

35.6

-1

36.0

-2

36.0

GPU (standard)

fp16

3477.6

3453.8

4656.6

11588.0

23192.5

34780.5

fp32

3406.5

3540.4

5168.1

12115.0

22817.3

34932.3

GPU (optimal)

fp16

572.9

550.8

653.9

1777.6

1686.3

3463.9

fp32

996.8

815.4

2732.5

4544.7

1747.3

6292.0

  1. 1The long- and short-term components both affect the conductance of the same device so that the addition W+F does not need to be explicitly performed.
  2. 2The input-weight multiplication is computed via I = G â‹… U, which consumes power during the read operation. This power consumption is, however, already accounted for by the Decay term, which requires the application of a constant bias Vbias to control the Λ parameter. Here, the worst-case scenario is assumed (Vbias = 0.6V applied to all synapses). The electrical current due to this voltage can be read out, giving the result of the input-weight multiplication. A crossbar array configuration is necessary to enable the full vector-matrix multiplication of all inputs with all synapses.
Back to article page