PerGB

Cheapest GPUs to run gpt-oss 120B locally

gpt-oss 120B needs about 72 GB of VRAM at Q4 with an 8K-token context (125 GB at Q8). A mixture of experts: 5.1B active. The cheapest way to hold it right now: 4 × Tesla P40 for $1,083, about 34–50 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
4 × Tesla P4096 GB20.0 GB spare34–50 tokens/s$1,083the 4 cheapest listings1,000 WCUDA 8.0 to 12.9 (not CUDA 13)
3 × Instinct MI50 32 GB96 GB21.0 GB spare70–119 tokens/s$2,027the 3 cheapest listings900 WNot on ROCm's support list (older ROCm or Vulkan)
3 × Radeon Pro W680096 GB21.0 GB spare35–60 tokens/s$2,650the 3 cheapest listings750 WROCm (on AMD's support list)
4 × GeForce RTX 2080 Ti 22 GB88 GB12.0 GB spare60–90 tokens/s$2,679the 4 cheapest listings1,000 WCUDA 10.0 and newer
3 × Tesla V100 PCIe 32 GB96 GB21.0 GB spare87–131 tokens/s$2,929the 3 cheapest listings750 WCUDA 9.0 to 12.9 (not CUDA 13)
4 × GeForce RTX 3080 20 GB80 GB4.0 GB spare88–133 tokens/s$3,364the 4 cheapest listings1,280 Wover a 15 A circuitCUDA 11.1 and newer
4 × Arc Pro B60 24 GB96 GB20.0 GB spare31–53 tokens/s$3,407the 4 cheapest listings800 WNo CUDA (oneAPI / Vulkan)
3 × Tesla V100S PCIe96 GB21.0 GB spare110–165 tokens/s$3,436the 3 cheapest listings750 WCUDA 9.0 to 12.9 (not CUDA 13)
4 × Radeon RX 7900 XT80 GB4.0 GB spare54–93 tokens/s$3,489the 4 cheapest listings1,260 Wover a 15 A circuitROCm (on AMD's support list)
4 × Titan RTX96 GB20.0 GB spare65–98 tokens/s$3,840the 4 cheapest listings1,120 WCUDA 10.0 and newer
4 × Quadro RTX 600096 GB20.0 GB spare65–98 tokens/s$3,970the 4 cheapest listings1,040 WCUDA 10.0 and newer
2 × Quadro RTX 800096 GB22.0 GB spare65–98 tokens/s$4,000the 2 cheapest listings520 WCUDA 10.0 and newer
3 × Instinct MI10096 GB21.0 GB spare83–143 tokens/s$4,029the 3 cheapest listings900 WROCm (on AMD's support list)
2 × Arc Pro B60 48 GB96 GB22.0 GB spare31–53 tokens/s$4,104the 2 cheapest listings400 WNo CUDA (oneAPI / Vulkan)
2 × A16128 GB54.0 GB spare78–116 tokens/s$4,439the 2 cheapest listings500 WCUDA 11.1 and newer

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.