RTX 30 Series — five cards, 8 GB to 24 GB

Which of the five

Cheapest first. Read down until the memory column holds what you need, then stop — the cards above it are faster at the same job, not capable of a different one.

The trap in the middle of that table

Our RTX 3080 is the original 10 GB card, the one with 8,704 CUDA cores. NVIDIA later shipped a 12 GB revision with 8,960, and a great many listings elsewhere quote the 12 GB figure beside the cheaper card. Ten gigabytes is not twelve: it is the difference between holding a thirteen-billion-parameter model at eight-bit precision and not holding it. If you are sizing against a number you read somewhere else, size against ours.

The other one to watch is the step from the 3080 Ti to the 3090. It is the largest capacity jump on the ladder — twelve gigabytes to twenty-four — and it is the only one of the five that clears 24 GB. If you need that, nothing cheaper on this page will do, and the alternatives are on the whole catalogue, priced.

The specifications, as NVIDIA publishes them

Second-generation ray-tracing cores, third-generation tensor cores, PCIe Gen 4 throughout. Figures from NVIDIA’s own GeForce comparison pages, cited in the source of this page.

Which machine they go in, and what link they get there

Every RTX 30 card is a PCIe Gen 4 part. Three of the four platforms that take them are Gen 3, and one is Gen 4.

For this generation the slot rarely decides anything: a card that holds its own scene or its own weights barely touches the bus once it is loaded, and these are not cards anyone streams enormous training batches to. It is recorded here because it is true and because we publish it for every card, not because it should change your mind.

What actually fits: the 24 GB card at the top of the ladder

Read that table down a rung for each step down the ladder. The 3080 Ti holds twelve gigabytes, the 3080 ten, the 3070 Ti and 3070 eight apiece — so a seven-to-eight-billion-parameter model at eight-bit precision is roughly where the bottom of this page lands, and everything larger wants the 3090 or a different shelf entirely.

What they cost, with the machines they go in

The card is one line on the invoice and the machine is another. Both halves below come from one reading of the order catalogue, so the price and the specification beside it always belong to each other.

Bandwidth is unmetered in both directions with no egress charge, one IPv4 address is included, there is no contract and one month is the default term. For every other card we fit, priced the same way, read the whole catalogue, priced. For machines racked and ready today, /instant. For why we still rent hardware everyone else retired, /value.