“NVIDIA's Vera Rubin chip architecture is specifically designed for the pre-fill stage of inference, featuring very little high-bandwidth memory to reduce cost.”