Researchers introduce PTXBench, a benchmark designed to evaluate and adapt large language models (LLMs) for GPU kernel optimization using architecture-specific PTX. The benchmark assesses functional correctness, target instruction execution at runtime, and performance speedup over existing libraries across GEMM and attention workloads on H100 and B200 GPUs. Results reveal that while architecture-specific PTX shows promise, success rates decline significantly on complex attention backward workloads.