experiments.md

Results

MNIST

1. StochasticLutSimple (LUT6)

• Training results
  • accuracy: 90.24%, epoch_size: 120, mini_batch_size: 32, lut_binarize: 1, 4 layers
• Synthesis results
  • LEs: 7835 (slack: +89.169ns)
  • clock frequency: 10MHz (constraint- 100ns)
• accuracy : 87%, epoch_size: 8 (7897 LEs)

2. StochasticLutCnn (LUT6)

• Training results:

  • accuracy 89.3 %, epoch size: 8, mini_batch_size: 32, lut_binarize: 1

• Synthesis results[ERROR-> generated verilog is incorrect]:

  • LEs: (slack: )
  • clock frequency: 10MHz ( constraint: 100ns )

3. StochasticLutSimple (LUT 4):

• Training results (ERROR -> Not able to converge/oscillating accuracies)
  • accuracy:15%, epoch_size: 120, mini_batch_size: 32, lut_binarize: 1 , 4 layers (608 LEs with slack +94ns )
  • accuracy: 64 %, epoch_size: 10, mini_batch_size: 120, lut_binarize: 1, 7 layers
• Synthesis results:
  • LEs:( slack: )
  • clock frequenccy: 10 MHz (constraint- 100ns)

4. StochasticLutCnn (LUT4)

• Training results - accuracy:
• Synthesis results: - LEs: - clock frequenccy: 10 MHz (constraint- 100ns)

CIFAR10

1.