Best hardware for
Requires ~20 GB VRAM (Q4 + KV cache). Ranked by estimated decode tok/s across 90 GPUs.
24 GB VRAM · $6,950
24 GB VRAM
32 GB VRAM · $3,900
GPUs are ranked by estimated decode tok/s for CodeLlama 34B. Higher tok/s means faster text generation during inference. Rankings combine community benchmarks, lab measurements, and spec-based estimates when real data isn't available.
"Fits" vs "VRAM tight": GPUs marked "Fits" have enough VRAM for the full model with headroom for KV cache growth. "VRAM tight" GPUs may work with shorter context lengths but could run out of memory with long conversations.
Multi-GPU setup: Use multiple GPUs to split the model across VRAM pools. Two RTX 4090s (48 GB total) can run larger models that don't fit on a single card.
Quantization: Reduce model precision with Q4 or Q8 quantization to fit in less VRAM. Quality trade-off is usually minimal for most use cases.
Cloud alternatives: Consider GPU cloud rental for occasional use instead of purchasing hardware for models that require extensive VRAM.
16 GB VRAM · $1,199
16 GB VRAM · $945
16 GB VRAM
48 GB VRAM
24 GB VRAM
80 GB VRAM
24 GB VRAM
16 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
24 GB VRAM
48 GB VRAM
80 GB VRAM
24 GB VRAM
24 GB VRAM
16 GB VRAM
12 GB VRAM
16 GB VRAM · $1,250
12 GB VRAM
12 GB VRAM
12 GB VRAM
12 GB VRAM