“GPU architecture is no longer optimal for AI workloads, as its fixed ratio of compute, memory, and memory bandwidth is out of step with the needs of advanced models.”