GPUs for training a large language model
Datacentre parts with the memory to hold optimiser state, not just weights.
The rule this page selects on
Datacentre tier, 80GB or more of memory, cheapest on-demand rate first.
Training holds weights, gradients and optimiser state at once — commonly several times the memory inference needs for the same model. 80GB is where a 13B full fine-tune stops spilling; below it you are into sharding, which is a different question from which card to rent.
20 priced parts carry no memory figure on record and are excluded rather than assumed to qualify. Unknown is not a yes.
Every part that meets it
On-demand list rates, read from each provider’s own page within 21 days · open any part for every provider selling it
What this ranking does not know
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.