benchmark_master_cifar100_subset_2026-02-28.txt

ResNet-50 Speed Benchmark (Backprop vs Predictive vs Circadian)
---------------------------------------------------------------
Device: cuda
Dataset: cifar100 (root=data, size=96, batch=32, augmentation=True, subset_train=20000, subset_test=5000)

BackpropResNet50
  epochs=12, cross_entropy=1.7144, acc=0.678
  training: 244.58s total, 981.3 samples/s, 17.69 ms/step
  inference: mean=21.33 ms, p95=23.03 ms, 1500.4 samples/s
  params: total=23,712,932, trainable=204,900

PredictiveCodingResNet50
  epochs=12, cross_entropy=1.1175, acc=0.692
  training: 248.65s total, 965.2 samples/s, 18.50 ms/step
  inference: mean=17.88 ms, p95=20.77 ms, 1789.7 samples/s
  params: total=24,333,348, trainable=825,316
  vs backprop: acc_delta=+0.014, train_samples_per_second_delta=-16.1
  energy=0.0006

CircadianPredictiveCodingResNet50
  epochs=12, cross_entropy=1.1082, acc=0.685
  training: 274.53s total, 874.2 samples/s, 22.03 ms/step
  inference: mean=21.68 ms, p95=23.27 ms, 1475.9 samples/s
  params: total=24,333,348, trainable=825,316
  vs backprop: acc_delta=+0.007, train_samples_per_second_delta=-107.0
  energy=0.0005
  circadian sleep: hidden=384->384, splits=0, prunes=0, rollbacks=0