PerGB

Cheapest GPUs to run Gemma 3 27B locally

Gemma 3 27B needs about 21 GB of VRAM at Q4 with an 8K-token context (34 GB at Q8). The cheapest way to hold it right now: 2 × Tesla P100 PCIe 12 GB for $257.00, about 17–25 tokens/s.

The cheapest setups right now

Live on eBay
SetupVRAMSpeed (est.)Price nowPowerSoftware
2 × Tesla P100 PCIe 12 GB24 GB0.7 GB spare17–25 tokens/s$257.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
Tesla P4024 GB1.7 GB spare10–16 tokens/s$270.00250 WCUDA 8.0 to 12.9 (not CUDA 13)
2 × Tesla P100 PCIe 16 GB32 GB8.7 GB spare22–33 tokens/s$345.00the 2 cheapest listings500 WCUDA 8.0 to 12.9 (not CUDA 13)
2 × Radeon VII32 GB8.7 GB spare22–37 tokens/s$503.00the 2 cheapest listings600 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Instinct MI50 16 GB32 GB8.7 GB spare22–37 tokens/s$538.62the 2 cheapest listings600 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Tesla V100 PCIe 12 GB24 GB0.7 GB spare27–41 tokens/s$543.85the 2 cheapest listings500 WCUDA 9.0 to 12.9 (not CUDA 13)
2 × Radeon RX 6700 XT24 GB0.7 GB spare8.1–14 tokens/s$580.00the 2 cheapest listings460 WNot on ROCm's support list (older ROCm or Vulkan)
2 × GeForce RTX 2060 12 GB24 GB0.7 GB spare10–15 tokens/s$589.99the 2 cheapest listings320 WCUDA 10.0 and newer
Instinct MI50 32 GB32 GB9.7 GB spare22–37 tokens/s$633.68300 WNot on ROCm's support list (older ROCm or Vulkan)
2 × Tesla V100 PCIe 16 GB32 GB8.7 GB spare27–41 tokens/s$659.00the 2 cheapest listings500 WCUDA 9.0 to 12.9 (not CUDA 13)
2 × GeForce RTX 3060 12 GB24 GB0.7 GB spare11–16 tokens/s$698.99the 2 cheapest listings340 WCUDA 11.1 and newer
2 × Radeon Pro VII32 GB8.7 GB spare22–37 tokens/s$725.00the 2 cheapest listings500 WNot on ROCm's support list (older ROCm or Vulkan)
4 × GeForce RTX 2070 Super32 GB6.7 GB spare13–20 tokens/s$739.97the 4 cheapest listings860 WCUDA 10.0 and newer
4 × Quadro RTX 400032 GB6.7 GB spare13–19 tokens/s$744.96the 4 cheapest listings640 WCUDA 10.0 and newer
4 × GeForce RTX 207032 GB6.7 GB spare13–20 tokens/s$749.47the 4 cheapest listings700 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.