Set a budget and a model size, and we rank the discrete GPUs you can buy for it — using the same fit engine as the “will it run?” calculator. Prices are dated and sourced; no store links, no affiliate links.
Prefer to describe hardware you already own? Use the “will it run?” calculator.
Judged with a representative catalog model for this class: DeepSeek-R1-Distill-Qwen-14B (14.8B), best GREEN quant at 8,192 tokens of context. Speeds are estimates, not measurements.
How this list is ranked (reproducible from the numbers below)
Assumes a single card plus 32 GB of system RAM (the same assumption the per-GPU pages use). Two-card combos are out of scope for now — see the per-GPU pages for multi-GPU notes.
NVIDIA Data Center (used)
Estimated 31–42 tok/s (estimate)
NVIDIA Data Center (used)
Estimated 12–16 tok/s (estimate)
AMD Radeon RX 7000 (RDNA 3)
Estimated 33–44 tok/s (estimate)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 32–43 tok/s (estimate)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 35–47 tok/s (estimate)
NVIDIA Data Center (used)
Estimated 41–55 tok/s (estimate)
NVIDIA Data Center (used)
Estimated 51–68 tok/s (estimate)
Intel Arc (Alchemist)
Estimated 32–42 tok/s (estimate)
AMD Radeon RX 6000 (RDNA 2)
Estimated 29–39 tok/s (estimate)
AMD Radeon RX 7000 (RDNA 3)
Estimated 35–47 tok/s (estimate)
AMD Radeon RX 6000 (RDNA 2)
Estimated 29–39 tok/s (estimate)
AMD Radeon RX 9000 (RDNA 4)
Estimated 36–48 tok/s (estimate)
NVIDIA GeForce RTX 50 (Blackwell)
Estimated 51–68 tok/s (estimate)
AMD Radeon RX 9000 (RDNA 4)
Estimated 36–48 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 38–51 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 42–56 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 41–54 tok/s (estimate)
NVIDIA GeForce RTX 50 (Blackwell)
Estimated 25–34 tok/s (estimate)
NVIDIA GeForce RTX 50 (Blackwell)
Estimated 54–72 tok/s (estimate)
NVIDIA RTX Workstation (Ampere)
Estimated 25–34 tok/s (estimate)
AMD Radeon RX 7000 (RDNA 3)
Estimated 28–37 tok/s (estimate)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 20–27 tok/s (estimate)
Intel Arc (Battlemage)
Estimated 26–34 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 29–38 tok/s (estimate)
NVIDIA GeForce RTX 50 (Blackwell)
Estimated 38–51 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 29–38 tok/s (estimate)
NVIDIA GeForce RTX 40 (Ada Lovelace)
Estimated 29–38 tok/s (estimate)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 3–5 tok/s (estimate) (with offload)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 4–6 tok/s (estimate) (with offload)
NVIDIA GeForce RTX 30 (Ampere)
Estimated 3–5 tok/s (estimate) (with offload)
What’s excluded, and why