代理評測
FluxBench Puts Chip Agents Through the Full RTL-to-GDS Flow, With Performance Gaps of Up to 86% on the Same Model Architecture
FluxBench uses standardized prompts, tools, and technology libraries to test agents across the full flow from RTL generation through synthesis, placement and routing, and ECO, rather than evaluating only one-shot code output. The authors report that different agent frameworks using the same base model can exhibit a performance gap of 86.27%, although the commercial EDA experiments are limited in scope and the paper was previously withdrawn and resubmitted.
A study that recently drew attention in the technical community introduces FluxBench, an attempt to determine whether AI coding agents can truly complete a chip’s RTL-to-GDS workflow rather than merely generate a plausible-looking piece of Verilog. The benchmark standardizes prompts, tool environments, and the technology library, covering RTL generation, iterative fixes based on compiler feedback, logic synthesis, placement and routing, and Engineering Change Order automation. It includes both open-source toolchains and a closed commercial EDA flow using PicoRV32 as a case study.
The central finding is that using the same base model does not cause agent performance to converge. The authors report task-performance gaps of up to 86.27% between different agent system architectures. Among systems with similar completion rates, Token ROI—which measures effective design improvement relative to token usage and execution cost—varied by as much as 105.92×. FluxEDA achieved an end-to-end score of up to 97.94 on PicoRV32 and was claimed to outperform Claude Code equipped with a domain Skill by as much as 8.39×. This supports an important engineering conclusion: simply adding prompt documentation or domain knowledge cannot replace state management, error recovery, tool-feedback parsing, and cross-stage context transfer.
However, these multipliers cannot be directly extrapolated to advanced chips. The paper covers only a limited set of frameworks and models, along with a single closed industrial flow, and some tools and environments are difficult for external researchers to reproduce fully. The end-to-end score and Token ROI are also metrics designed by the authors. In addition, the arXiv record shows that v2 was withdrawn and v3 was re-uploaded on July 23, so any cited figures should be checked against the latest version. The next things worth tracking are a public harness, larger and more varied RTL designs and process nodes, and whether agent-generated results can pass formal verification, achieve timing closure, and satisfy power constraints—not merely complete the tool flow.