The answer — one pick, priced
Get the
NVIDIA RTX 3090 (Renewed)
No budget ceiling
The consensus best-value GPU for local AI — 24GB at ~$60/GB, roughly half the RTX 5090's cost per gigabyte, and every ML framework supports its Ampere architecture. 7 sources agree.
Some links earn us a commission — it never changes the pick. Winners are chosen before payouts are checked.
Fit ledger — need → measured evidence
■ The catch — co-equal billing, always
It's old silicon, bought used. Ampere's 3rd-gen Tensor Cores deliver just 71 FP16 TFLOPS — under a third of the RTX 5090's 209.5 — so training iterations run markedly slower, and you're buying on the renewed/used market with the reliability risk that carries.
Dealbreaker? Runner-up №1 — a new RTX 4070 Ti SUPER — for 8GB less VRAM ↓
■ Every product has a catch. Verdicts that hide it are how bad buys happen.
Runners-up — if your needs differ
Adjacent verdicts — same method
Traps — this verdict avoids
Training eats 2-4x the VRAM
A model that fits in VRAM for inference often won't fit for training — optimizer states (Adam uses 2x the model parameters), gradients, and activations push training memory to 2-4x inference. Size against training, not inference.
MSRP is fiction
The RTX 5090's $1,999 MSRP is nominal — real board-partner listings run ~$4,250, and the discontinued 4090 trades at $3,000+ where in stock. Budget the street price, never the MSRP.
12 GB is a dead end
Cards with 12 GB or less are effectively unusable for real training. 16 GB is the floor — and even that hits a wall at 13B+ parameters without aggressive quantization.
Provenance — 7 sources, dated
Winners are picked before affiliate payouts are checked, from the full research dossier at knowledgelib.io. Prices, stock and listings are re-verified monthly by an automated pipeline. Some links earn us a commission — it never changes the pick, and we say so here rather than in a footer you'd never read.