It'll help, but GPU crunch isn't caused by people running 6-8bit inference on a single card, but by all the large scale pre-training + fine-tuning runs.