Recent advances in large language models (LLMs) have spawned coding agents capable of tackling complex engineering tasks such as register‑transfer level (RTL) design and optimization. Existing RTL benchmarks mainly assess functional correctness and performance‑power‑area (PPA) metrics, leaving the agents' ability to achieve timing closure largely unevaluated. To fill this gap we introduce TicTacBench, a benchmark designed to evaluate coding agents' RTL‑level timing‑closure capability under post‑place‑and‑route (post‑PnR) conditions. The suite comprises 30 diverse tasks, each supplying a sub‑optimal RTL design, realistic timing constraints, functional equivalence checks, and timing reports. Over 300 runs across eight state‑of‑the‑art LLMs reveal that even the best agent closes only 53.3% of the tasks, incurring an average 7.18% degradation in area‑delay product (ADP) and an 8.83% improvement in energy‑delay‑squared product (EDDP). By categorising common failure modes we identify why agents miss timing closure, and we propose TicTacSkill, a method that guides agents through standard timing‑closure procedures, boosting the timing‑closure rate by roughly 9%. The findings suggest that while coding agents have made substantial strides in RTL design, their timing‑closure capability still has considerable room for improvement.
Review