How fast is Intel Arc B580 12GB for local AI workloads?
Intel Arc B580 12GB has a source-attributed result of 28.0 x RT on Whisper transcription (batch 1, 0-token context, INT8; runtime not documented). This is a reference report, not an independently verified lab result. A 70B Q4_K_M artifact needs roughly 40GB for weights alone; smaller memory configurations require explicit offload or model splitting and do not establish full-GPU residency.