GPUs for image generation

Diffusion models fit in far less than a language model.

16GB or more of memory, most memory first.

The one page here not ordered by price, and the reason is the workload. A diffusion checkpoint is a fraction of the size of a language model, so memory stops being a pass/fail constraint early — and past that point every extra gigabyte buys a bigger batch, which buys throughput. For serving a language model the opposite holds: once the weights fit, more memory does nothing and the cheapest card wins. Same floor, genuinely different answer.

20 priced parts carry no memory figure on record and are excluded rather than assumed to qualify. Unknown is not a yes.

It orders by hourly rate, which is not the same as cost. A card at half the price that takes three times as long is dearer, and we hold no benchmark that would tell you which — so no page here claims a cost per unit of work. Nor does it know about FP64 throughput, interconnect bandwidth or encoder blocks. If those decide your choice, this page has narrowed the field and not made the decision.