PlaidML Deep Learning Framework Benchmarks With OpenCL On NVIDIA & AMD GPUs

While the ROCm 2.0 stack was playing well with this OpenCL deep learning framework where as many other deep learning frameworks are catered towards NVIDIA's CUDA interfaces, the training performance in particular was very low out of the Radeon GPUs at least for VGG16 and VGG19. Even the Radeon RX Vega 64 was coming in shy of the GTX 1060 and GTX 980.

With the inference workloads on PlaidML, the Radeon GPUs were more competitive though still coming in a bit shy of expectations with for the VGG16 network the RX Vega 64 matching the GTX 1070.

With training for the IMDB LSTM model, the Radeon performance was very strong with the Radeon RX Vega cards competing with the GeForce GTX 1080 series and RTX 2070.

Mobilenet training was another case of the Radeon performance coming up short.