Find the cheapest GPU for the job.

The order

Cheapest total per hour. The price is for the whole configuration — four cards at $0.40 each shows as $1.60.

Only what you can rent

Node sizes come from the offers themselves. If a provider rents a card ×1 and ×8, we won’t price ×3 — separate rentals don’t pool their memory.

The pills

• No FP8 in hardware runs in FP16 instead — about twice the memory shown, so the fit may not hold.
• No BF16 FP16 only. BF16 weights usually convert, but can overflow.
• No FlashAttention-2 falls back to attention that uses more memory than estimated.
• Pools over PCIe no NVLink, so cross-GPU traffic goes through host memory.
• Interconnect not recorded this card ships with and without NVLink and the provider doesn’t say which. Ask before a multi-card run.
• ROCm / Habana, not CUDA different runtime. CUDA kernels need a port.

What reorders the list

Only one thing: a card that can’t do your selected precision is ranked on memory it wouldn’t use, so it sorts below cards that can.

Nothing is ordered by speed. We publish no throughput figures because we measure none — so a cheap old card can still cost more for the same job.

Sources

Prices are scraped from each provider and trace to an archived copy of the page. Architecture and format support come from vendor documentation, maintained by hand — capability only, never performance.