Transformer inference performance is a hardware–software co-design problem. AI hardware accelerator choices determine the available compute throughput, data movement, and storage, while program choices determine how effectively prompt prefill and autoregressive decode use those resources.
In this homework, use your own AI agent to optimize the hardware and software together. Submit one AI hardware accelerator configuration and two assembly programs that run on the same hardware, produce correct Transformer outputs, and improve the combined simulated score. A hardware choice is valuable only when the programs can exploit it, and every program optimization must be evaluated on the shared hardware. The full assignment statement provides the exact workload, rules, starter package, and submission format.