Local AI inference relies more on VRAM capacity, memory bandwidth, and the peak compute of dedicated matrix operation units, which is fundamentally different from the rasterization or ray-tracing throughput tested in gaming scenarios.