Each module has almost double the tensor cores, 4-5x the memory bandwidth and has nvlink support for all units. Meaning layer offloading mode should fly. Power consumption (and heat) is the major drawback. These can pull like 2500 watts during generation. However id suspect limiting that to less than half would have marginal impact on processing or generation speeds. I've seen as low as 80 watt limit per card cause approximately less than 10% impact on speeds. 500w -> 160w savings for 10% performance impact, on my less than ideal configuration.