PerGB

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
SetupVRAMSpeed (est.)Price nowPowerSoftware
2 × Tesla P4048 GB0.0 GB spare4.0–6.1 tokens/s$540.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
4 × Tesla P100 PCIe 16 GB64 GB14.0 GB spare8.6–13 tokens/s$721.00the 4 cheapest listings1,000 WCUDA 8.0 to 12.9 (not CUDA 13)
4 × Radeon VII64 GB14.0 GB spare8.4–14 tokens/s$1,023the 4 cheapest listings1,200 WNot on ROCm's support list (older ROCm or Vulkan)
4 × Instinct MI50 16 GB64 GB14.0 GB spare8.4–14 tokens/s$1,087the 4 cheapest listings1,200 WNot on ROCm's support list (older ROCm or Vulkan)
4 × Tesla V100 PCIe 16 GB64 GB14.0 GB spare11–16 tokens/s$1,327the 4 cheapest listings1,000 WCUDA 9.0 to 12.9 (not CUDA 13)
2 × Instinct MI50 32 GB64 GB16.0 GB spare8.4–14 tokens/s$1,329the 2 cheapest listings600 WNot on ROCm's support list (older ROCm or Vulkan)
4 × Radeon Pro VII64 GB14.0 GB spare8.4–14 tokens/s$1,630the 4 cheapest listings1,000 WNot on ROCm's support list (older ROCm or Vulkan)
4 × Radeon RX 680064 GB14.0 GB spare4.2–7.2 tokens/s$1,658the 4 cheapest listings1,000 WNot on ROCm's support list (older ROCm or Vulkan)
4 × Radeon RX 7600 XT64 GB14.0 GB spare2.4–4.0 tokens/s$1,660the 4 cheapest listings760 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Arc Pro B60 24 GB48 GB0.0 GB spare3.7–6.4 tokens/s$1,699the 2 cheapest listings400 WNo CUDA (oneAPI / Vulkan)
2 × Radeon Pro W680064 GB16.0 GB spare4.2–7.2 tokens/s$1,760the 2 cheapest listings500 WROCm (on AMD's support list)
4 × Quadro RTX 500064 GB14.0 GB spare5.2–7.9 tokens/s$1,839the 4 cheapest listings920 WCUDA 10.0 and newer
2 × Titan RTX48 GB0.0 GB spare7.9–12 tokens/s$1,840the 2 cheapest listings560 WCUDA 10.0 and newer
4 × A264 GB14.0 GB spare2.3–3.5 tokens/s$1,858the 4 cheapest listings240 WCUDA 11.1 and newer
4 × Arc A770 16 GB64 GB14.0 GB spare4.6–7.9 tokens/s$1,872the 4 cheapest listings900 WNo CUDA (oneAPI / Vulkan)

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.