Table 3.

Summary of results

DatasetState of the artTrueNorth best accuracyTrueNorth 1 chip
ApproachAccuracyAccuracy#coresAccuracy#coresFPSmWFPS/W
CIFAR10CNN (11)91.73%89.32%3149283.41%40421249204.46108.6
CIFAR100CNN (34)65.43%65.48%3149255.64%40421526207.87343.7
SVHNCNN (34)98.08%97.46%3149296.66%40422526256.59849.9
GTSRBCNN (35)99.46%97.21%3149296.50%40421615200.68051.8
LOGO32CNN93.70%90.39%1360685.70%32361775171.710335.5
VADMLP (36)95.00%97.00%175895.42%423153926.159010.7
TIMIT Class.HGMM (37)83.30%82.18%880279.16%19432610142.618300.1
TIMIT FramesBLSTM (38)72.10%73.46%2003871.17%24762107165.912698.0
  • The network for LOGO32 was an internal implementation. BLSTM, bidirectional long short-term memory; CNN, convolutional neural network; FPS, frames/second; FPS/W, fames/second per Watt; HGMM, hierarchical Gaussian mixture model; MLP, multilayer perceptron. Accuracy of TrueNorth networks is shown in bold.