Which RTX 50 card for running models locally

You have decided on a PC with a discrete NVIDIA card for running language models at home, and the shop lists four cards from the current generation. The hardware guide already settled how to choose between machines: the model's file has to fit in the memory of the device that runs it, and the guide's rule of thumb keeps that file to three quarters of the memory so the text the model is working on has room beside it. This comparison applies that one rule to the four cards, from the figures NVIDIA publishes for each, and stops where the figures stop.

The table above is not written here. It is built, on every build of this site, from a signed record for each card, and every number in it is a row from NVIDIA's own specification page for that card, quoted and checked. The prose below reads the table; it does not add to it.

What we compare

Four GeForce RTX 50 cards: the RTX 5090, the RTX 5080, the RTX 5070 Ti and the RTX 5070. These four because they are the desktop cards NVIDIA describes on the three specification pages the hardware guide already cites, and because a reader choosing a card for models today is choosing among them. Nothing older is in the table, and nothing from another maker: a comparison across makers would need a common figure both publish the same way, and this one keeps to a family where the rows line up.

Each column is the card's record: eleven figures the family defines — memory, memory interface width, CUDA cores, AI TOPS, boost clock, total graphics power, required system power, length, width, slots and price — each one a quotation from the maker's page with the number and the unit read out of it. The date under a card's name is the day the record was last read against its page and signed. When one card is re-checked and another is not, the two dates differ, and the table says so rather than pretending to one date for all.

The table

Memory comes first because the guide's rule makes it the only figure that decides whether a model runs at all; the rest of the rows describe how the card runs a model that fits, and what the card asks of the machine around it. Read a column top to bottom and you have the card; read a row across and you have the difference.

The units are the maker's: GB of GDDR7 for memory, bits for the interface width, watts for the two power figures, millimetres for the dimensions. Where a cell reads "The maker's page states this varies by manufacturer", that is what NVIDIA's page prints — for the RTX 5070 Ti's length, width and slots it gives no number, because partners build that card to their own dimensions. Where the price row reads that the page states no price a machine can read, that is also what the page does: it prints a placeholder to a crawler, so no price is a figure here and none is guessed. Nothing in the table was measured by this site. Every figure is the maker's, and the maker can change it; that is why each column carries a date.

Where the difference shows

Memory is where the four separate, and it separates them into three tiers rather than four. NVIDIA's specifications page lists the GeForce RTX 5090's standard memory configuration as 32 GB GDDR7. NVIDIA's specifications page lists the GeForce RTX 5080's standard memory configuration as 16 GB GDDR7. NVIDIA's RTX 5070-family specifications page lists the GeForce RTX 5070 Ti's standard memory configuration as 16 GB GDDR7 and the RTX 5070's as 12 GB GDDR7, in one shared row. Apply the guide's three-quarters rule and the room for a model's file is 24 GB on the RTX 5090, 12 GB on the RTX 5080 and the RTX 5070 Ti alike, and 9 GB on the RTX 5070.

Set the published file sizes against those four numbers and the choice becomes concrete. Ollama's library lists qwen3:4b at 2.5GB, qwen3:8b at 5.2GB, qwen3:14b at 9.3GB and qwen3:32b at 20GB. Ollama's library lists gemma3:4b at 3.3GB, gemma3:12b at 8.1GB and gemma3:27b at 17GB. Ollama's library lists llama3.1:8b at 4.9GB and llama3.1:70b at 43GB. So: the 20 GB qwen3:32b and the 17 GB gemma3:27b fit the RTX 5090's 24 GB of room and nothing else in this table; the 9.3 GB qwen3:14b fits the 12 GB of room on the RTX 5080 and the RTX 5070 Ti but not the 9 GB on the RTX 5070; the 8.1 GB gemma3:12b and everything smaller fit all four; the 43 GB llama3.1:70b fits none of them, and no setting changes that.

The second figure the guide names is bandwidth, and between cards of one generation the interface width is the figure to compare for it. NVIDIA's specifications page lists the GeForce RTX 5090's memory interface width as 512-bit. NVIDIA's specifications page lists the GeForce RTX 5080's memory interface width as 256-bit. NVIDIA's RTX 5070-family specifications page lists the GeForce RTX 5070 Ti's memory interface width as 256-bit and the RTX 5070's as 192-bit, in one shared row. The RTX 5080 and the RTX 5070 Ti share both the memory size and the width, which is why the table shows them as near twins on the two rows that matter most for models; the RTX 5070 is narrower as well as smaller, and the RTX 5090 is twice as wide as the middle pair.

Power is where the cost of the top card shows before any price does. NVIDIA's specifications page lists the GeForce RTX 5090's total graphics power as 575 W. NVIDIA's specifications page lists the GeForce RTX 5090's required system power as 1000 W. NVIDIA's specifications page lists the GeForce RTX 5080's total graphics power as 360 W. NVIDIA's specifications page lists the GeForce RTX 5080's required system power as 850 W. NVIDIA's RTX 5070-family specifications page lists the GeForce RTX 5070 Ti's total graphics power as 300 W and the RTX 5070's as 250 W, in one shared row. NVIDIA's RTX 5070-family specifications page lists the GeForce RTX 5070 Ti's required system power as 750 W and the RTX 5070's as 650 W, in one shared row. The step from the RTX 5080 to the RTX 5090 doubles the memory and adds 215 W to the card's own draw and 150 W to the power supply NVIDIA asks for.

What it means for you

If the models you want to run are the ones up to about 8 GB — the 4B, 8B and 12B sizes in Ollama's library — any of the four cards holds them with room to spare, and the choice among the four is not about models at all. The RTX 5070's 9 GB of room covers that list; its narrower interface means the words come somewhat slower than from the cards above it, but they come.

If you want the 14B class, the RTX 5070 is the one to cross off: its 9 GB of room is under the 9.3 GB file, and the guide's rule exists precisely so that you do not plan on the last gigabyte. The RTX 5080 and the RTX 5070 Ti both hold it, and on the two rows that decide this — memory and interface width — the table cannot tell them apart. Choose between them on the figures the table does show that differ, the power draw, and on what you can find to buy.

If you want the 27B or 32B class on one card, only the RTX 5090 in this table has the memory, and it asks for a 1000 W power supply to go with it. Before you buy it, read the label on the supply already in your machine; the guide's Step 2 reads the card you have now, and this comparison adds the second thing to check.

Nothing here tells you what any of the four costs. NVIDIA's pages print no price a machine can read, so no price is a figure in the table, and a comparison that added one from somewhere else would no longer be built from the maker's own page.

The sources

Three pages at NVIDIA, one per card or pair of cards, as the cards themselves cite them: the RTX 5090's page, the RTX 5080's page, and the RTX 5070-family page that lists the RTX 5070 Ti and the RTX 5070 side by side in shared rows. Each column's figures were read from its page and signed on the date under the card's name, and that date is the card's own — a card re-checked later shows a later date without this comparison being rewritten. The model file sizes are Ollama's library pages for qwen3, gemma3 and llama3.1, the same pages the hardware guide cites, quoted at the sizes the library serves by default. The list that follows, under its own heading, has one entry per card: the card's name leading to its own page, the address of the maker's page it was read from, and the date it was last verified.