Curious how the memory bandwidth actually holds up under sustained inference loads compared to the spec sheet numbers.