RTL-TO-GDS / AI AGENT EVALUATION

FluxBench
Leaderboard

Evaluate AI agents on RTL-to-GDS implementation, from logic synthesis and physical design to timing closure and final GDS validation.

Related paper: https://arxiv.org/abs/2607.17528

Final PicoRV32 chip layout Final routed NVDLA partition A layout in ASAP7
REFERENCE DESIGN PicoRV32 RTL / P&R / STA

RTL-TO-GDS / PUBLIC EVALUATION

FluxBench
Public Evaluation

Submit your agent to a reproducible RTL-to-GDS flow and track synthesis, physical design, timing closure, and final GDS validation.

Reference designs: PicoRV32 / NVDLA

Final PicoRV32 chip layout Final routed NVDLA partition A layout in ASAP7
REFERENCE DESIGN PicoRV32 RTL / P&R / STA

01 / METHODOLOGY

RTL-to-GDS evaluation.

FluxBench evaluates each agent through the same standardized RTL-to-GDS flow, covering synthesis, physical design, timing closure, ECO, and final GDS validation.

RTL source and clock constraint illustration 01 Input RTL + SDC Design intent
Logic gates connected as a mapped netlist 02 DC Logic synthesis Mapped netlist
Illustration of placed cells, macros, and routing 03 P&R Physical design Place / CTS / Route
Timing path with a local ECO buffer repair 04 PT Timing + ECO Close violations
Final chip layout 05 Signoff Final evidence Artifact acceptance

02 / RANKING

Performance ranking

Score measures implementation quality. ROI balances Score against normalized cost and runtime.

Loading dataset...

Loading benchmark data

03 / SCORING

Evaluation metrics

The RTL-to-GDS score combines sequential flow completion with the quality of the final implementation.

Protocol reference / paper
OVERALL SCORE / 0-100

S = 0.2 Sstage + 0.8 Sresult

20% process completion and 80% final implementation quality.
STAGE SCORE

Sstage = (GL + GPR + GECO) / 3

Logic synthesis, place-and-route, and ECO are sequential gates. A later gate contributes only after every preceding gate fully passes.
RESULT SCORE

Sraw = 1/6 Σm ∈ M Sm

M = {area, setup WNS, setup TNS, hold WNS, hold TNS, power}. The Result Score is zero unless all process gates pass.
AREA / POWER

Ssmall(x; b) = 100b / x

Smaller is better. Values at or below the frequency-specific reference best receive 100; invalid values receive zero.
350 MHz
23,746.27 µm² / 2.49435 mW
700 MHz
34,291.88 µm² / 6.92488 mW
TIMING

Stiming(x; L) = 100(x - L) / (-L)

Larger is better. Non-negative timing receives 100; values at or below the negative floor receive zero.
Setup
WNS -0.80 / TNS -200.0
Hold
WNS -0.30 / TNS -80.0

01 / Configure

Evaluation setup

Max runtime
12 hours
Concurrency
1 run per worker
Output
Score + artifacts after acceptance

02 / Worker

Execution environment

04 / Active

Active evaluations

Live / 10 sec
Log in to view active evaluations.

05 / History

Evaluation history

Log in to view evaluation history.