All tracked GPUs

From consumer RTX cards to professional workstation GPUs, AMD RDNA3, and every Apple Silicon chip from M1 to M5. This is the full list of hardware CanItRun tracks for local LLM inference. VRAM capacity determines which models fit natively; memory bandwidth determines how fast they run.

Click any GPU to see a full compatibility breakdown: which open-weight models run natively in VRAM, which require CPU offload, and estimated tokens per second for each. For Apple Silicon, unified memory means the full RAM pool is available: an M3 Max with 128GB can run models that require a multi-GPU NVIDIA setup.