# AI Chip Design Results: How to Evaluate an RTL Optimization Claim

OpenAI's chip-design results make a case for testing AI-assisted RTL work. A useful evaluation must hold features, area budgets and implementation targets steady.

By Paras Madan | 2026-11-18


OpenAI's [November 18 chip-design report](https://openai.com/index/advancing-chip-design-with-our-models/) combines production work with research evaluations. It reports up to 14% improvement in power, performance and area (PPA) against heavily optimized AMD GPU IP baselines, with some changes incorporated into a design sent for manufacturing. Separately, it evaluates models optimizing or generating PicoRV32 and Ibex processor cores. Those are different kinds of evidence, not interchangeable performance claims.

The RISC-V tests reject designs that fail simulation checks or exceed sixteen times the original area. Throughput combines instructions per cycle with a Yosys-estimated clock frequency. Scores are normalized against the best observed result under the constraints, so a score of 0.7 is not a 70% production improvement. OpenAI also notes that specification-driven designs can omit features, potentially explaining some advantage over editing existing RTL.

For a hardware team, the practical next step is a carefully bounded experiment. Here is the evaluation plan I would use, rather than treating the research chart as a prediction for a particular chip.

## Freeze the configuration before asking for improvements

RTL, or register-transfer-level code, describes the circuit's behavior. Before changing it, save the source revision, configuration parameters, tool versions, target technology and test workload. Identify which parts of the design may change and which interfaces must remain compatible.

This matters even with a familiar open core. The [PicoRV32 README](https://github.com/YosysHQ/picorv32) documents multiple instruction-set configurations, optional interrupts and selectable register-file implementations. A smaller result after disabling a required feature is a product change, not a like-for-like optimization.

I would put the requirements in a short acceptance document beside the baseline. Include supported instructions, interrupt behavior, reset behavior and software compatibility. Ask the model to propose a narrow change and explain which requirement each test exercises. Keep that explanation as a review aid, not as evidence that the circuit is correct.

## Make the target technology explicit

The [Ibex register-file documentation](https://ibex-core.readthedocs.io/en/latest/03_reference/register_file.html) illustrates why a generic area comparison can mislead. It offers flip-flop, FPGA and latch-based implementations. The latch-based version cannot be simulated with Verilator and needs a technology-specific clock-gating cell. The FPGA implementation instead uses appropriate memory resources where available.

Decide which implementation belongs in your baseline before running the experiment. Otherwise, an apparent model improvement may actually come from changing the hardware target or selecting a different implementation family.

For every candidate, retain the synthesis settings and generated reports. If the candidate needs a tool or primitive your build does not support, record that as an integration cost rather than hiding it behind the headline result.

## Separate correctness from optimization

I would use two gates. First, establish that the change preserves the required behavior using the team's verification flow. Then compare implementation results. Do not let a faster candidate skip the correctness gate because it looks promising on a few example workloads.

Keep the final evaluation workload outside the model's optimization loop. During development, give it actionable failures from a separate test set. Record rejected candidates too, including time spent diagnosing them. Engineering effort is part of the cost of finding a usable improvement.

## Report the trade-off, not a single winning number

Choose an area ceiling appropriate to the actual project, alongside timing and power requirements. Measure the baseline and candidate through the same flow and distinguish early synthesis estimates from later implementation results.

My decision record would show the retained features, verification outcome, area change, timing change and unresolved risks. A candidate that improves one dimension while exceeding another limit should remain a trade-off for review, not become an automatic success.

That is the useful way to take these results seriously: give the model room to propose changes, but make the engineering contract precise enough that a higher score cannot quietly mean a different chip.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn research results and engineering tradeoffs into clear technical stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
