OpenAI has demonstrated how artificial intelligence can accelerate semiconductor engineering after its AI models helped AMD achieve up to 14% improvement in power, performance and area (PPA) on parts of its flagship GPU intellectual property. The collaboration delivered the reported improvements in less than five months, building on hardware designs that AMD engineers had already spent years optimizing. OpenAI estimates that the work represents approximately a year of additional chip-design progress compressed into a much shorter period.

The development is significant because AI is moving beyond writing software and generating content into the specialized engineering processes used to build the hardware powering artificial intelligence itself. OpenAI’s models rewrote sections of AMD’s chip-design code, while industry-standard electronic design automation tools verified the changes. AMD has already incorporated some of the optimizations into a design sent for manufacturing, with additional changes added to production design files for future versions.

OpenAI AI Models Deliver Up to 14% PPA Improvement for AMD

OpenAI announced the results on November 18, 2026, as part of a broader update on using its AI models for advanced chip design.

The collaboration focused on AMD’s flagship GPU intellectual property, or IP. Semiconductor companies use IP blocks as reusable building components for processors. These blocks can handle calculations, data movement, memory operations and other functions inside a chip.

According to OpenAI, its models identified improvements of up to 14% in PPA metrics compared with heavily optimized existing designs.

PPA stands for:

  • Power: The electricity required for a chip or circuit to operate.
  • Performance: How quickly the hardware can execute operations.
  • Area: The amount of physical silicon occupied by the circuit.

Improving these three metrics is one of the most difficult challenges in semiconductor engineering. Engineers often spend years refining a design to extract relatively small improvements without affecting its functionality.

OpenAI says its models found additional opportunities by rewriting the SystemVerilog code describing AMD’s circuits.

Key Details of the AMD Collaboration

MetricReported Result
Maximum PPA improvementUp to 14%
Development periodLess than five months
Existing engineering baselineAlready optimized by AMD engineers
Verification toolsSynopsys electronic design automation tools
Manufacturing statusSome changes included in a design sent for manufacturing
Additional optimizationsAdded to production design files for future versions

The 14% figure represents the maximum reported improvement, not an average across an entire GPU. OpenAI has not disclosed the exact distribution of gains across power, performance and chip area or identified the specific GPU product involved.

How OpenAI Used AI to Redesign Chip Circuits

The most important part of the collaboration was the use of AI to modify SystemVerilog RTL.

RTL stands for Register Transfer Level. It is a hardware-description approach used by engineers to specify how digital circuits operate and how information moves between different parts of a processor.

SystemVerilog is widely used in semiconductor development to describe complex hardware designs before they are manufactured.

OpenAI’s models analyzed existing RTL code and proposed alternative implementations of circuit blocks.

For example, a model could examine a circuit responsible for performing mathematical calculations and identify a different way to implement the same function using fewer resources or less time.

The AI-generated designs were then checked using Synopsys tools.

These tools helped verify that the rewritten circuits preserved their intended behavior and evaluate the trade-offs involving power consumption, performance and physical area.

This verification step is essential because a circuit that appears more efficient but produces incorrect results cannot be used in a commercial processor.

OpenAI says its models demonstrated reasoning capabilities comparable to those of an expert microarchitect when working with RTL.

AMD Has Already Incorporated Some AI-Generated Improvements

The collaboration has progressed beyond experimental simulations.

OpenAI says AMD has already included some of the AI-generated changes in a recent design submitted for manufacturing. Other optimizations have been incorporated into production design files for future versions.

This is important because semiconductor companies cannot rely solely on improvements that look promising in software simulations.

A design must pass extensive verification and manufacturing processes before it can become part of a commercial chip.

However, the fact that some changes have entered a manufacturing-bound design does not mean that the complete 14% improvement has already been demonstrated in a commercially available GPU.

AMD’s manufacturing process and final product performance will determine how much of the reported optimization translates into real-world benefits.

OpenAI Also Improves Its Own Jalapeño AI Chip

The AMD collaboration is not the only example of OpenAI applying its models to semiconductor engineering.

The company has also used similar techniques to improve Jalapeño, its first custom inference chip.

Inference chips are designed to execute trained AI models and generate outputs. They are particularly important for operating large language models because serving millions of user requests requires substantial computing resources.

OpenAI reported several improvements to Jalapeño:

Jalapeño ComponentReported Improvement
Warp engine standard-cell areaApproximately 13% reduction
Matrix unit standard-cell areaApproximately 10% reduction
Multiplier cell areaApproximately 15% reduction
Multiplier latencyReduced from two cycles to one

Standard-cell area refers to the amount of silicon occupied by the basic logic components used to implement a circuit.

OpenAI says these optimizations were verified during Jalapeño’s actual silicon bring-up, meaning the improvements were tested on manufactured hardware rather than being limited to simulations.

This provides an additional example of AI-assisted chip optimization progressing into physical hardware.

OpenAI Tests AI Models on Open-Source RISC-V Processors

OpenAI also evaluated how well its models could optimize existing processor designs.

The company used two open-source 32-bit RISC-V processor cores:

  • PicoRV32
  • Ibex, which is used in the OpenTitan project.

The models received hardware descriptions and were asked to improve processor throughput.

The evaluation involved approximately 9,881 scored runs, with a token budget of around 112,000 tokens per attempt.

OpenAI compared several models, including GPT-6.2 Astra, GPT-6 Astra and GPT-6.1 Sol.

GPT-6.2 Astra achieved the strongest average results across both processor benchmarks.

An interesting finding was that all tested models performed better when asked to design a processor from a written specification rather than optimize an existing implementation.

However, OpenAI acknowledged an important limitation: a model starting from a specification could omit features present in the original processor. Therefore, higher benchmark scores do not automatically mean that the redesigned processor is functionally equivalent to the original.

This distinction matters because chip design requires strict correctness, not simply higher performance scores.

OpenAI Partners With Synopsys to Expand AI-Assisted Chip Design

OpenAI is also working with Synopsys, a major provider of semiconductor design and verification software.

The partnership aims to combine OpenAI’s AI models with Synopsys’ electronic design automation tools.

One planned development is GPT-Synopsys, a specialized model designed to interact with Synopsys software and assist chip designers with complex engineering workflows.

The objective is to make AI-assisted semiconductor development more accessible to engineers working on different types of processors and hardware.

Synopsys tools are widely used across the semiconductor industry for designing, simulating and verifying integrated circuits.

Integrating AI into these workflows could help engineers explore more design alternatives, automate repetitive tasks and identify optimization opportunities faster.

Why This Development Matters for the AI Industry

The semiconductor industry is under increasing pressure to deliver more powerful and energy-efficient processors.

Demand for GPUs is growing as companies expand AI training, inference, cloud computing and data-center infrastructure.

However, improving chip performance is becoming increasingly difficult. Advanced processors contain billions of transistors, and even small architectural changes can create complicated engineering challenges.

AI-assisted design could help address this problem by exploring a much larger number of possible circuit implementations than human engineers can manually evaluate.

For AMD, the collaboration could provide additional opportunities to improve processor efficiency and strengthen its position in the AI hardware market.

For OpenAI, the development demonstrates another application of its models beyond conventional software development.

The company is not only building AI systems that run on advanced chips but also using AI to help design the hardware itself.

The Bigger Picture

The collaboration between OpenAI and AMD highlights an emerging cycle in the technology industry: AI is increasingly being used to improve the hardware required to run AI.

More efficient circuits could help reduce power consumption, improve processing performance and increase the amount of computing available from a given amount of silicon.

These improvements could eventually benefit data centers, AI developers and companies purchasing advanced processors.

However, the reported 14% improvement should be understood as a maximum result on parts of AMD’s flagship GPU IP, not a guaranteed 14% increase in the performance of an entire graphics processor.

The broader significance is that AI models are beginning to contribute to highly specialized engineering tasks that previously required extensive manual experimentation.

Looking Ahead

The next important milestone will be seeing how much of the AI-generated optimization translates into manufactured AMD processors and future commercial products. The company has already included some changes in a design sent for manufacturing, but further information will be needed to establish the impact on real-world power consumption and performance.

For OpenAI, the collaboration could become a foundation for expanding AI-assisted hardware engineering through partnerships with semiconductor design software companies such as Synopsys. If these systems become reliable across more complex chip-design workflows, AI could significantly change how processors are developed, tested and optimized in the coming years.


SEO Title: OpenAI Helps AMD Achieve 14% GPU Design Improvement Using AI

Meta Description: OpenAI’s AI models helped AMD achieve up to 14% PPA improvement on flagship GPU IP in under five months through AI-assisted chip design.

Suggested URL Slug: openai-amd-ai-chip-design-14-percent-ppa-improvement

Focus Keyword: OpenAI AMD chip design

Secondary Keywords: AMD GPU optimization, OpenAI AI models, 14% PPA improvement, SystemVerilog RTL, Synopsys, AI chip design, Jalapeño AI chip, semiconductor engineering

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.