The catch is volatility: one CUDA process reclaims the VRAM, and your swap just evaporates.