Eh, this is way overblown IMO. The product page claims this is for training, and as long as you crank your batch size high enough you will not run into memory bandwidth constraints.

I've finetuned diffusion models streaming from an SSD without noticeable speed penalty at high enough batchsize.