fp16 is usable for training and inference, so the Vega would have been good for both. I care about inference performance too, anyway. CPU is only a (much slowe

fp16 is usable for training and inference, so the Vega would have been good for both. I care about inference performance too, anyway. CPU is only a (much slower) fallback if the OpenCL drivers are broken.