Old Titan X Pascal hits 25 tok/s

Old datacenter-grade Pascal hardware still punches above its weight for local inference, which matters as more developers look to repurpose aging GPUs rather than buy new.

• –25 tok/s generation on decade-old hardware is genuinely usable for background coding agents or overnight batch tasks
• –llama.cpp's CPU+GPU offloading means even cards with limited VRAM can contribute meaningfully to inference throughput
• –The 5x speedup over CPU-only (6 tok/s → 25 tok/s) shows the GPU floor is low but real
• –This is a practical benchmark for the "basement server" crowd increasingly running local AI pipelines