Radeon ROCm 2.0 OpenCL Benchmarks With Linux 5.0 On Ubuntu 18.10 vs. NVIDIA's Linux Driver
With LeelaChessZero (lczero) that uses deep learning for chess, the Radeon performance remains quite low with the RX Vega 64 coming in just shy of the GeForce GTX 1060. As from previous benchmarks, this is an area where the software could be better optimized for Radeon GPUs and/or a mix of areas for improvement too within the ROCm software stack.
With the PlaidML machine learning benchmark, the Radeon RX Vega 64 was aligned with the GeForce GTX 1070 Ti rather than being up with the GeForce GTX 1080 or better. But it will be interesting to see where the Radeon VII fits into this scheme soon enough.
With the ResNet 50 model and FP16 precision, the RX Vega 64 was coming in short of the GTX 1070 Ti while the RX 590 was also coming in well below the GTX 1060.
With the IMDB LSTM network and avoiding FP16, the Radeon Vega performance was quite competitive with the RX Vega 56 even coming just in front of the GeForce GTX 1080.
For Mobilenet the RX Vega 56 was less competitive but at least the RX Vega 64 still stepped out in front of the GeForce GTX 1080.
Like with the IMDB LSTM training, the inference performance was also very good with Vega on ROCm 2.0.
The VGG16 inference performance meanwhile was not quite as competitive with Vega but not as bad as some of the tests.