[D] Much lower wattage if dual GPU used for training

I have a dual GPU setup, 4080 16GB & 3090 24GB. With my toy example of training Bert Large on the Yelp review dataset, I consistently get around \~4% slower training on dual GPU than on the 4080 only. I've expected the 2 GPUs will be able to work nicely together as they have very similar CUDA core counts (4080 **9,728** ,3090 **10,496**). The other observation the 3090 pulls only around 270w in dual training while training only on the …

bert dataset example gpu gpus machinelearning review setup training will work yelp