Cheapest GPUs to run Llama 3.3 70B locally
Llama 3.3 70B needs about 46 GB of VRAM at Q4 with an 8K-token context (78 GB at Q8). The cheapest way to hold it right now: 2 × Tesla P40 for $540.00, about 4.0–6.1 tokens/s.
The cheapest setups right now
Live on eBay| Setup | VRAM | Speed (est.) | Price now | Power | Software |
|---|---|---|---|---|---|
| 2 × Tesla P40 | 48 GB0.0 GB spare | 4.0–6.1 tokens/s | $540.00the 2 cheapest listings | 500 W | CUDA 8.0 to 12.9 (not CUDA 13) |
| 4 × Tesla P100 PCIe 16 GB | 64 GB14.0 GB spare | 8.6–13 tokens/s | $721.00the 4 cheapest listings | 1,000 W | CUDA 8.0 to 12.9 (not CUDA 13) |
| 4 × Radeon VII | 64 GB14.0 GB spare | 8.4–14 tokens/s | $1,023the 4 cheapest listings | 1,200 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Instinct MI50 16 GB | 64 GB14.0 GB spare | 8.4–14 tokens/s | $1,087the 4 cheapest listings | 1,200 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Tesla V100 PCIe 16 GB | 64 GB14.0 GB spare | 11–16 tokens/s | $1,327the 4 cheapest listings | 1,000 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 2 × Instinct MI50 32 GB | 64 GB16.0 GB spare | 8.4–14 tokens/s | $1,329the 2 cheapest listings | 600 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Radeon Pro VII | 64 GB14.0 GB spare | 8.4–14 tokens/s | $1,630the 4 cheapest listings | 1,000 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Radeon RX 6800 | 64 GB14.0 GB spare | 4.2–7.2 tokens/s | $1,658the 4 cheapest listings | 1,000 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 4 × Radeon RX 7600 XT | 64 GB14.0 GB spare | 2.4–4.0 tokens/s | $1,660the 4 cheapest listings | 760 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 2 × Arc Pro B60 24 GB | 48 GB0.0 GB spare | 3.7–6.4 tokens/s | $1,699the 2 cheapest listings | 400 W | No CUDA (oneAPI / Vulkan) |
| 2 × Radeon Pro W6800 | 64 GB16.0 GB spare | 4.2–7.2 tokens/s | $1,760the 2 cheapest listings | 500 W | ROCm (on AMD's support list) |
| 4 × Quadro RTX 5000 | 64 GB14.0 GB spare | 5.2–7.9 tokens/s | $1,839the 4 cheapest listings | 920 W | CUDA 10.0 and newer |
| 2 × Titan RTX | 48 GB0.0 GB spare | 7.9–12 tokens/s | $1,840the 2 cheapest listings | 560 W | CUDA 10.0 and newer |
| 4 × A2 | 64 GB14.0 GB spare | 2.3–3.5 tokens/s | $1,858the 4 cheapest listings | 240 W | CUDA 11.1 and newer |
| 4 × Arc A770 16 GB | 64 GB14.0 GB spare | 4.6–7.9 tokens/s | $1,872the 4 cheapest listings | 900 W | No CUDA (oneAPI / Vulkan) |
Per question: scaled from the model's reported speed on an RTX PRO 6000 by each card's compute and memory bandwidth, within about half either way. Licenses, scores and sources: decision models.
A set of cards is priced at that many of the cheapest live listings, shipping included. Speed: an estimate from memory bandwidth for one user chatting, within a third either way.
How it works
The memory a model needs is its weights at the quantisation you pick, plus the attention cache for the context and a little for working buffers. Q4 is what most people run at home, close to full quality at a third of the size. Each card keeps about 1 GB for itself, so two 24 GB cards hold 46 GB.
Writing each token reads every active weight once, so the speed is about the memory bandwidth divided by the model's size, times what the software gets out of the card (less off CUDA). Cards sharing a model layer by layer take turns, so a set runs about as fast as one card would. Treat it as a rough figure for one person chatting; a long prompt takes longer to read in.
Power is the cards' rated power. Over about 1,440 W with the rest of the PC, a set needs more than one 15 A circuit (or a 240 V one), and older cards often need a power cable adapter, which each card's page mentions.