Performance Level 4

CUDA support

Convolutional input layer

2.0

Conv and pooling layers are working

Performance Level 3

Performance measurement:
Hardware:
Laptop (ASUS FX753VE): Intel i7-7700HQ / NVIDIA GTX 1050Ti (Mobile)
Desktop: Intel i5 6600k / Palit Jetstream NVIDIA GTX 970

Configuration:
Neural Net: (784 | 30 | 10) fully connected
Precision: double
Training data: MNIST 60k
MiniBatchSize: 8
MiniBatchesPerEpoch: 7.500
Epochs: 5

Performance Level 3:
Parallelized

1. Feed forward and error calculation by subdividing the mini batch and processing partial mini batches parallel
2. Gradient calculation and parameter adjustment by doing it parallel for the biases and weights on each layer

Redesigned linear algebra without memory allocation during training
x64 RelWithDebInfo (/MD /Zi /O2 /Ob1 /DNDEBUG)

Results:
Laptop: N/A
Desktop: 15.2s

Performance Level 2

Performance measurement:
Hardware:
Laptop (ASUS FX753VE): Intel i7-7700HQ / NVIDIA GTX 1050Ti (Mobile)
Desktop: Intel i5 6600k / Palit Jetstream NVIDIA GTX 970

Configuration:
Neural Net: (784 | 30 | 10) fully connected
Precision: double
Training data: MNIST 60k
MiniBatchSize: 8
MiniBatchesPerEpoch: 7.500
Epochs: 5

Performance Level 2:
Redesigned linear algebra without memory allocation during training
x64 RelWithDebInfo (/MD /Zi /O2 /Ob1 /DNDEBUG)

Results:
Laptop: N/A
Desktop: 35.8s

Version 0.1 - Performance Level 0/1

Performance measurement:
Hardware:
Laptop (ASUS FX753VE): Intel i7-7700HQ / NVIDIA GTX 1050Ti (Mobile)
Desktop: Intel i5 6600k / Palit Jetstream NVIDIA GTX 970

Configuration:
Neural Net: (784 | 30 | 10) fully connected
Precision: double
Training data: MNIST 60k
MiniBatchSize: 8
MiniBatchesPerEpoch: 7.500
Epochs: 5

Performance Level 0:
x64 RelWithDebInfo (/MD /Zi /O2 /Ob1 /DNDEBUG)

Results:
Laptop: 136.1s
Desktop: 123.3s

Performance Level 1:
Parallel minibatch error calculation
x64 RelWithDebInfo (/MD /Zi /O2 /Ob1 /DNDEBUG)

Results:
Laptop: 117.5s
Desktop: 101.7s