PerGB

Cheapest GPUs to run gpt-oss 20B locally

gpt-oss 20B needs about 14 GB of VRAM at Q4 with an 8K-token context (23 GB at Q8). A mixture of experts: 3.6B of its 21B parameters work on each token. The cheapest way to hold it right now: Tesla P100 PCIe 16 GB for $170.00, about 101–151 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
Tesla P100 PCIe 16 GB16 GB1.4 GB spare101–151 tokens/s$170.00250 WCUDA 8.0 to 12.9 (not CUDA 13)
Radeon VII16 GB1.4 GB spare99–169 tokens/s$249.00300 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Tesla P100 PCIe 12 GB24 GB8.4 GB spare75–113 tokens/s$257.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
Instinct MI50 16 GB16 GB1.4 GB spare99–169 tokens/s$266.00300 WNot on ROCm's support list (older ROCm or Vulkan)
Tesla P4024 GB9.4 GB spare48–71 tokens/s$270.00250 WCUDA 8.0 to 12.9 (not CUDA 13)
Tesla V100 PCIe 16 GB16 GB1.4 GB spare124–186 tokens/s$325.00250 WCUDA 9.0 to 12.9 (not CUDA 13)
2 × GeForce RTX 2070 Super16 GB0.4 GB spare62–92 tokens/s$349.99the 2 cheapest listings430 WCUDA 10.0 and newer
Radeon Pro VII16 GB1.4 GB spare99–169 tokens/s$355.00250 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Quadro RTX 400016 GB0.4 GB spare57–86 tokens/s$359.99the 2 cheapest listings320 WCUDA 10.0 and newer
2 × GeForce RTX 207016 GB0.4 GB spare62–92 tokens/s$365.50the 2 cheapest listings350 WCUDA 10.0 and newer
2 × GeForce RTX 208016 GB0.4 GB spare62–92 tokens/s$375.00the 2 cheapest listings430 WCUDA 10.0 and newer
Radeon RX 680016 GB1.4 GB spare49–84 tokens/s$380.00250 WNot on ROCm's support list (older ROCm or Vulkan)
2 × GeForce RTX 2060 Super16 GB0.4 GB spare62–92 tokens/s$387.51the 2 cheapest listings350 WCUDA 10.0 and newer
Radeon RX 7600 XT16 GB1.4 GB spare28–48 tokens/s$399.99190 WNot on ROCm's support list (older ROCm or Vulkan)
2 × GeForce RTX 2060 8 GB16 GB0.4 GB spare46–69 tokens/s$405.68the 2 cheapest listings320 WCUDA 10.0 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.