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| Setup | VRAM | Speed (est.) | Price now | Power | Software |
|---|---|---|---|---|---|
| 4 × Tesla P40 | 96 GB20.0 GB spare | 34–50 tokens/s | $1,083the 4 cheapest listings | 1,000 W | CUDA 8.0 to 12.9 (not CUDA 13) |
| 3 × Instinct MI50 32 GB | 96 GB21.0 GB spare | 70–119 tokens/s | $2,027the 3 cheapest listings | 900 W | Not on ROCm's support list (older ROCm or Vulkan) |
| 3 × Radeon Pro W6800 | 96 GB21.0 GB spare | 35–60 tokens/s | $2,650the 3 cheapest listings | 750 W | ROCm (on AMD's support list) |
| 4 × GeForce RTX 2080 Ti 22 GB | 88 GB12.0 GB spare | 60–90 tokens/s | $2,679the 4 cheapest listings | 1,000 W | CUDA 10.0 and newer |
| 3 × Tesla V100 PCIe 32 GB | 96 GB21.0 GB spare | 87–131 tokens/s | $2,929the 3 cheapest listings | 750 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 4 × GeForce RTX 3080 20 GB | 80 GB4.0 GB spare | 88–133 tokens/s | $3,364the 4 cheapest listings | 1,280 Wover a 15 A circuit | CUDA 11.1 and newer |
| 4 × Arc Pro B60 24 GB | 96 GB20.0 GB spare | 31–53 tokens/s | $3,407the 4 cheapest listings | 800 W | No CUDA (oneAPI / Vulkan) |
| 3 × Tesla V100S PCIe | 96 GB21.0 GB spare | 110–165 tokens/s | $3,436the 3 cheapest listings | 750 W | CUDA 9.0 to 12.9 (not CUDA 13) |
| 4 × Radeon RX 7900 XT | 80 GB4.0 GB spare | 54–93 tokens/s | $3,489the 4 cheapest listings | 1,260 Wover a 15 A circuit | ROCm (on AMD's support list) |
| 4 × Titan RTX | 96 GB20.0 GB spare | 65–98 tokens/s | $3,840the 4 cheapest listings | 1,120 W | CUDA 10.0 and newer |
| 4 × Quadro RTX 6000 | 96 GB20.0 GB spare | 65–98 tokens/s | $3,970the 4 cheapest listings | 1,040 W | CUDA 10.0 and newer |
| 2 × Quadro RTX 8000 | 96 GB22.0 GB spare | 65–98 tokens/s | $4,000the 2 cheapest listings | 520 W | CUDA 10.0 and newer |
| 3 × Instinct MI100 | 96 GB21.0 GB spare | 83–143 tokens/s | $4,029the 3 cheapest listings | 900 W | ROCm (on AMD's support list) |
| 2 × Arc Pro B60 48 GB | 96 GB22.0 GB spare | 31–53 tokens/s | $4,104the 2 cheapest listings | 400 W | No CUDA (oneAPI / Vulkan) |
| 2 × A16 | 128 GB54.0 GB spare | 78–116 tokens/s | $4,439the 2 cheapest listings | 500 W | CUDA 11.1 and newer |
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.