Every shipping NVIDIA, AMD and Intel GPU from Pascal (2016) to Blackwell/Rubin (2025–26),
with VRAM type, capacity, bandwidth and current street prices. Prices are live from
GPUDojo's tracker as of 2026-10-03 where listed; ranges are cross-checked estimates.
Filter and sort below — click a column header to sort.
Reading the table
CUDA EOL cliff: Pascal (GTX 10 / P-series) and Volta (V100, Titan V) lost CUDA 12 compiler targets — legacy PyTorch builds only. Turing (T4 / RTX 20) is the oldest fully-supported CUDA generation.1
ROCm cliff: official ROCm starts at RDNA 2 (RX 6000). Polaris/Vega/MI50/MI100 are legacy builds; MI100 has no INT8 at all.
Intel: works via oneAPI / ipex-llm / vLLM-Intel — functional but a step behind CUDA/ROCm in plug-and-play.
Bandwidth > FLOPS for LLM inference. Model fit rule of thumb: ~2 GB per 4B params at 8-bit + 20% KV headroom.