The NVIDIA A40 GPU has been benchmarked on Lambda Scalar servers, comparing its performance to previous generation flagship GPUs such as the V100 and RTX series. The A40's 48 GB of GDDR6 memory provides consistent performance improvement across various deep learning models, including ConvNet and language models like BERT and TransformerXL. The GPU scales near perfectly with multi-GPU training, making it suitable for server use cases where passive cooling is sufficient. Its performance is comparable to its workstation peer, the A6000, but with a lower power consumption and no active cooling required.