ServicesBlogPricingContactContact Now
AI4 min read

OpenAI Models Accelerate Chip Design with AMD and RISC‑V

OpenAI Models Accelerate Chip Design with AMD and RISC‑V

Introduction

The semiconductor industry is increasingly turning to generative AI to tackle the intricate tasks of hardware design. Recent collaborations between OpenAI and leading silicon teams illustrate how frontier language models can contribute concrete improvements in power, performance, and area (PPA) for both custom inference chips and third‑party IP. This article examines the practical outcomes of these efforts, from the optimization of the Jalapeño compute blocks to measurable gains on AMD’s flagship GPU and CPU designs, and explores how open‑source RISC‑V cores are being evaluated for AI‑assisted enhancements.

Design Improvements on the Jalapeño Custom Chip

OpenAI began integrating GPT‑o3 mini into the Jalapeño development workflow in early 2025. Even at that stage the model identified non‑trivial opportunities to reduce area while preserving functionality. In the vector unit, which handles parallel compute, the model suggested modifications to floating‑point adders, multipliers, conversion functions, reduction logic, and the special‑function unit. By re‑organizing logic across pipeline stages, the proposals allowed hardware sharing and less intermediate storage.

Similar gains were observed in the matrix unit, an arithmetic block central to AI workloads. A multiplier redesign achieved roughly 15% less cell area during synthesis and cut latency from two clock cycles to a single cycle. An adder redesign also saved about 15% of cell area while keeping its two‑cycle latency. Further refinements to dot‑product circuits and pipeline organization added additional savings across the block.

Across the entire Jalapeño effort, the team estimates that model‑proposed changes reduced standard‑cell area by approximately 13% in the vector unit and दु 10% in the matrix unit. These optimizations were validated in silicon during the chip’s first fabrication run, confirming that AI‑generated modifications are not only theoretical but also manufacturable.

Collaboration with AMD and Real‑World Gains

OpenAI’s partnership with AMD focuses on applying the same AI‑driven methodology to third‑party hardware. The joint work has produced up to 14% PPA improvement relative to heavily optimized baselines on AMD’s flagship GPU IP. The model generated new SystemVerilog RTL descriptions that AMD engineers incorporated into a recent design sent for manufacturing and added to production design files for future silicon releases.

AMD also verified single‑digit percentage gains in PPA across key blocks of its upcoming next‑generation flagship CPU. The changes were derived from model‑written RTL, and each rewrite was checked using Synopsys electronic design automation (EDA) tools to ensure behavioral preservation and evaluate power, performance, (++?) area trade‑offs. The speed of iteration is notable: roughly a year of progress achieved in less than five months, a pace that OpenAI describes as unprecedented in the chip design domain.

OpenAI is also working with Synopsys to integrate frontier intelligence directly into the EDA tools that engineers already use. This effort aims to ground capability development in trusted evaluators, allowing designers to leverage AI suggestions while maintaining the rigor required for production‑grade verification.

Open‑Source RISC‑V Core Evaluation

To assess how model generations evolve over time, OpenAI tested each model on two open‑source 32‑bit RISC‑V cores: PicoRV32, a compact core for FPGAs, and OpenTitan’s Ibex processor used in a security chip. The evaluation task was to increase instruction throughput while respecting area constraints. Designs that failed verification or exceeded 16 times the original core’s area received a zero score.

For cores that passed, the team measured instructions per clock cycle on hidden workloads, took the geometric mean, and multiplied by the maximum clock frequency estimated with Yosys. The best observed throughput under the same constraints was normalized to a score of 1. compris. All five models scored higher when they designed a core from written requirements than when they attempted to improve an existing core.

With a comparable token budget of about 112,000 tokens per run, GPT‑6.2 Astra achieved the highest average score on both cores. GPT‑6.1 Sol also outperformed GPT‑5.6 Sol across the board. On the PicoRV32 core, GPT‑6.2 Astra’s mean score was roughly double that of GPT‑6 Astra at the same token budget. On the Ibex core, GPT‑6.2 Astra’s score continued to climb as the token budget increased, while GPT‑6 Astra and GPT‑6.1 Sol were closer together.

What Engineers Should Do Now

The demonstrated success stories provide a clear roadmap for practitioners. First, integrate AI assistance early in the design flow, starting with exploratory tasks such as logic restructuring and pipeline re‑balancing. Second, validate AI‑generated RTL using established EDA verification tools to ensure functional correctness and PPA trade‑offs. Third, establish a feedback loop where human experts review model suggestions, incorporate domain knowledge, and re‑run the model for iterative refinement. Fourth, document the changes and share them with downstream teams, as AMD has done by feeding model‑derived RTL into production design files. Finally, monitor token usage and set realistic expectations: the current models deliver gains within a comparable computational budget but still require human oversight for safety and compliance.

Grivyonx Expert Analysis

OpenAI’s progress in silicon design underscores a broader trend: AI is moving from concept to concrete engineering value. For security‑focused organizations, the ability to rapidly prototype and verify hardware changes is a strategic advantage. Grivyonx’s DevSecOps automation platform can help embed these AI‑driven workflows into existing CI/CD pipelines, ensuring that any model‑generated hardware changes are automatically validated against security policies before reaching production. Leveraging cloud‑native AI services for chip design also reduces on‑premise compute overhead, allowing teams to focus on higher‑level architectural decisions.

Conclusion

As AI models continue to refine their understanding of hardware description languages, the industry can expect even larger PPA improvements across a broader spectrum of silicon products. Grivyonx’s AI automation and cloud security services are well‑positioned to help organizations integrate these advances safely and efficiently into their development cycles.

  • #AI chip design
  • #OpenAI
  • #AMD
  • #RISC-V
  • #EDA tools
  • #GPU IP
  • #PPA optimization
Gourav Rajput

Gourav Rajput

Founder of Grivyonx Technologies