RTLScout Bench
Leaderboard
Convergence
Designs
Result vs. cost
Mean cost reached against mean API spend per run, at list prices from each run's recorded token usage. Runtime per run is in the table above.
Every run
The individual runs behind each row, with the evaluation and agent step at which the best design appeared. Failed evaluations are designs that did not pass the testbench or the flow.
Benchmark and method
Why area × delay is the score
Every design is scored by its area-delay product (ADP): the cell area of the synthesized netlist in µm² times its critical-path delay in ps. The critical path bounds the clock period, and the testbench fixes how many cycles the job takes, so a design's throughput is proportional to 1 / delay. Its throughput per unit silicon is therefore proportional to 1 / ADP: halving the ADP means the same die area delivers twice the compute, whether you get there by a smaller circuit, a faster one, or both. That is what makes ADP a meaningful single number rather than an arbitrary blend of two metrics.
It also keeps the optimization honest. Neither axis can be bought with the other at face value: a design that shrinks area by 20 % but lengthens the critical path by 25 % scores worse, and one that speeds up by duplicating hardware only wins if the speed-up exceeds the extra area. Because the cycle count is pinned by the cycle-accurate testbench, the agent cannot trade latency for area either; it has to find genuinely better logic.
The design
The score
The agent loop
Cost accounting
Dollars per run come from each run's recorded token usage (input and output tokens, plus cache writes and reads where a provider reports them) at the list prices given in the per-entry notes below. Prices and token counts are stored with every run, so the figure can be recomputed.