Back to News
InnovationAI Understanding briefing

OpenAI models deliver up to 14% performance‑power‑area gains on AMD chips

OpenAI says its frontier models have generated a year’s worth of chip‑design optimisation for AMD’s flagship GPU and CPU, achieving up to 14% PPA improvement in under five months.

4 min readRead the primary source
Source-provided image accompanying OpenAI models deliver up to 14% performance‑power‑area gains on AMD chips
Primary-source documentSource recorded
Publisher
openai.com
Source type
Primary document — an official announcement, paper, filing, or first-party page we read directly.
ContextUnderstand this in 60 seconds

Key terms

Generative AI
AI systems that produce new content such as text, images, audio, video, or code.
Benchmark
A standardized test or dataset used to measure and compare model performance.
Embedding
A numeric vector representation that captures semantic meaning of text, images, or other data.

What happened

OpenAI reported that its latest generative models, including GPT‑6.2 Astra, were used to rewrite SystemVerilog RTL for AMD’s next‑generation GPU and CPU designs, delivering up to 14% power‑performance‑area (PPA) gains versus heavily optimised baselines. The company also detailed how earlier models such as GPT‑4o contributed to the design of OpenAI’s own custom inference chip, Jalapeño, achieving roughly 13% area reduction in the vector unit and 10% in the matrix unit, improvements that were silicon‑proven during bring‑up. In tests on two open‑source 32‑bit RISC‑V cores (PicoRV32 and Ibex), GPT‑6.2 Astra achieved the highest average throughput scores, roughly doubling the score of earlier GPT‑6 models on PicoRV32 and outperforming GPT‑6.1 Sol and GPT‑5.6 Sol on both cores. The models performed these optimisations within a token budget of about 112,000 per run, using Synopsys EDA tools to verify functional correctness and evaluate trade‑offs.

OpenAI’s blog post outlines a series of chip‑design experiments where its generative models were tasked with improving the RTL of both OpenAI’s custom inference chip, Jalapeño, and AMD’s flagship GPU and CPU designs. The models explored alternative logic implementations, pipeline restructuring, and component sharing, leading to measurable reductions in standard‑cell area and latency. For Jalapeño, the vector unit saw a 13% area reduction and the matrix unit a 10% reduction, with specific multiplier and adder redesigns achieving up to 15% cell‑area savings.

In collaboration with AMD, the models rewrote SystemVerilog code for the GPU IP, achieving up to a 14% improvement in power‑performance‑area (PPA) relative to heavily optimised baselines. AMD has already incorporated some of these AI‑generated changes into a design sent for manufacturing and added others to future production files, indicating a move from experimental to production use.

The evaluation on open‑source RISC‑V cores involved giving each model the original hardware description and a token budget (~112,000 tokens) to optimise throughput. GPT‑6.2 Astra achieved the highest average scores, roughly doubling the performance of earlier GPT‑6 variants on the PicoRV32 core and outperforming GPT‑6.1 Sol and GPT‑5.6 Sol on both PicoRV32 and Ibex. The models also performed better when designing a core from written specifications rather than merely tweaking existing designs, suggesting strength in higher‑level architectural synthesis.

All model‑generated RTL changes were verified using Synopsys EDA tools to ensure functional correctness and to assess PPA trade‑offs. The partnership with Synopsys includes training a dedicated model, GPT‑Synopsys, to operate within the Synopsys toolchain, further AI capabilities into standard chip‑design workflows.

Source details: openai.com

Why it matters

The reported gains illustrate a concrete, production‑level impact of large‑scale AI models on semiconductor engineering, a field traditionally dominated by human experts and lengthy design cycles. By automating RTL rewrites and microarchitectural trade‑off analysis, OpenAI’s models can compress what would normally take months or years of engineering effort into weeks, potentially accelerating time‑to‑market for new chips and reducing development costs. For AMD, integrating AI‑generated optimisations into its flagship GPU and upcoming CPU could translate into more competitive products with better performance per watt, a critical metric in data‑center and edge‑computing markets. The partnership with Synopsys further signals a broader industry shift toward directly into electronic design automation (EDA) workflows, which may reshape talent requirements and toolchains across the chip design ecosystem. However, the results are currently demonstrated on specific designs and under controlled token‑budget constraints; broader applicability, long‑term reliability, and the economics of scaling such AI‑assisted design pipelines remain open questions.

The ability of AI models to produce silicon‑proven RTL changes demonstrates a shift from AI‑assisted design assistance to AI‑driven design creation, potentially reducing the time and expertise required for chip optimisation.

For semiconductor manufacturers, AI‑generated improvements could lower development costs, accelerate product cycles, and enhance competitive positioning, especially in performance‑critical markets like data centers and AI accelerators.

The collaboration with Synopsys suggests that AI integration may become a standard feature of EDA suites, influencing the broader ecosystem of chip design tools and possibly redefining the skill set needed for future chip engineers.

The reported token‑budget constraints and the need for extensive verification highlight that while promising, the approach still requires careful validation and may not yet replace human expertise for all design aspects.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

System Requirements:
Best ArchitecturePure RAGRecommended pattern
Hallucination RiskVery LowGrounding efficacy
Update Cost$0 (Vector sync)Ongoing maintenance
Core takeaway: Fine-tuning teaches models how to speak (form, style, syntax); RAG teaches models what to say (verifiable facts). Never use fine-tuning alone for factual memory.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Key indicators to monitor include AMD’s upcoming product releases that incorporate the AI‑generated RTL changes, any disclosed performance benchmarks of the next‑gen GPU and CPU, and whether other semiconductor firms adopt similar AI‑driven design flows. Follow the rollout of the GPT‑Synopsys model and its integration with Synopsys EDA tools, as well as any third‑party validation of the reported PPA improvements. Additionally, watch for industry commentary on the cost‑benefit balance of token‑budget‑driven optimisation versus traditional engineering effort, and any regulatory or security scrutiny of AI‑generated hardware designs.

AMD’s upcoming GPU and CPU product announcements to see if AI‑generated RTL changes are highlighted and how they affect performance.

Release notes or case studies from Synopsys regarding the GPT‑Synopsys model’s capabilities and adoption by other chip designers.

Industry reactions, including from other chipmakers like Intel, Nvidia, and TSMC, regarding the viability of AI‑driven RTL optimisation at scale.

Potential regulatory or security discussions about the provenance and verification of AI‑generated hardware designs, especially for critical infrastructure chips.

Related guides & quizzes

Found this useful?