Why can't this scale to run much larger models on CPU backed by flash with good access patterns?