The domain of face recognition has shown a dramatic development in the recent years. Most impressive results were encountered using deep convolutional networks. Driven by reasons of efficiency, we approached the problem using a simple architecture, namely a ResNet18 network with 11 million parameters. By using the Additive Margin Softmax loss function we show how the performance of a plain architecture with no changes can be improved close to the one of state-of-the-art models with an order of magnitude more parameters. We obtain a train accuracy of 96% and a validation accuracy of 87% and a ROC area close to 0.98 on the LFW dataset. We show that even such a simple architecture is appropriate for face recognition.