Multilayer perceptron (MLP)
The model below is a simple multilayer perceptron (MLP) with 3 layers.
class MLP(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super().__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.fc2 = nn.Linear(hidden_size, hidden_size)
self.fc3 = nn.Linear(hidden_size, output_size)
def forward(self, x):
x = nn.functional.relu(self.fc1(x))
x = nn.functional.relu(self.fc2(x))
return self.fc3(x)
mlp = MLP(256.0, 384.0, 128.0)
We demonstrate the model at a variety of batch sizes. The model has 295K parameters.
IA-840F: 3 big cores
Model | Batch size | Mean latency (μs) | 99th Percentile latency (μs) |
---|---|---|---|
mlp_b1 | 1 | 2.9 | 3.5 |
mlp_b4 | 4 | 3.1 | 3.7 |
mlp_b8 | 8 | 3.6 | 4.1 |
IA-420F: 6 small cores
Model | Batch size | Mean latency (μs) | 99th Percentile latency (μs) |
---|---|---|---|
mlp_b1 | 1 | 3.4 | 3.7 |
mlp_b4 | 4 | 3.7 | 4.0 |
mlp_b8 | 8 | 4.2 | 4.6 |