Table 1 Benchmarking results

From: Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip

GPU(3090)

 

Vanilla SNN

Dynamic SNN (This work)

Task

Acc(%)

Rest power (mW)

Total power (mW)

Latency (ms)

Spike counts (×106)

Acc(%)

Rest power (mW)

Total power (mW)

Latency (ms)

Spike counts (×106)

Gesture

82.3

30000

30078

24.7

1.2

92.0 (+9.7)

30000

30079

28.1

0.5 (−60.5%)

Gait-day

87.2

30000

30047

24.5

2.6

90.2 (+3.1)

30000

30048

27.5

1.0 (−60.8%)

Gait-night

85.5

30000

30049

21.5

3.5

91.0 (+5.6)

30000

30049

23.5

1.3 (−62.8%)

Speck

Gesture

81.0

0.42

9.5

<0.1

1.0

90.0 (+9.0)

0.42

3.8 (−60.0%)

<0.1

0.4 (−60.0%)

Gait-day

86.0

0.42

16.1

<0.1

2.9

90.0 (+4.0)

0.42

7.3 (−54.7%)

<0.1

1.2 (−58.6%)

Gait-night

86.0

0.42

46.8

<0.1

3.3

91.0 (+5.0)

0.42

12.3 (−73.7%)

<0.1

1.5 (−54.5%)

  1. On the Gesture30/Gait-day31/Gait-night32, we use exactly the same experimental settings, including input time window, network structure, training method, hyper-parameter settings, etc. We design a tiny network structure for these datasets, i.e., Input-32C3S1-32C3S2-32C3S1-AP4-32FC-Output. Note, AP4-average pooling with 4 × 4 pooling kernel size, nC3Sm-Conv layer with n output feature maps, 3 × 3 weight kernel size, and m stride size, kFC-Linear layer with k neurons. We first train the model on GPU (Nvidia RTX 3090) and then deploy the trained model to Speck (only five SNN cores are utilized). In the inference stage, we count the accuracy, energy consumption, and latency on both GPU and Speck. For the GPU, we set the batch size to 1, ran 1000 samples (all samples have an input time window of 540 ms), and counted their power and latency. In the whole process, we remove the consumption of the data loading process. For the Speck, we randomly sample 100 samples on each dataset as input to Speck chip part to evaluate power and latency, where the total power is composed of Logic and RAM power, and the latency of a single sample is defined as the difference between the timestamp of the output result and the last input event. Thanks to the Sinabs framework and the newly proposed spiking neuron model (please see “Methods” section), the model trained on GPU has little accuracy loss after being deployed on Speck. Note, since we use public event-based datasets, the power reported in this Table only involves the processor part of Speck and does not include the DVS camera part of Speck (please see Eq. (2)).
Back to article page