OpenAI Reports 14% Gains in AMD GPU Chip Design

OpenAI reports up to 14% PPA gains in selected AMD GPU circuits, with some AI-designed changes entering a manufacturing-bound chip design.

Published: November 19, 2026

By Ray Ashish

OpenAI Reports 14% Gains in AMD GPU Chip Design
OpenAI Reports 14% Gains in AMD GPU Chip Design

Artificial Intelligence is beginning to influence one of the most technically demanding stages of semiconductor development: designing the circuits inside advanced processors. OpenAI says its AI models have identified improvements of up to 14% in power, performance and area (PPA) measures on AMD’s flagship graphics processing unit (GPU) intellectual property. More significantly, AMD has reportedly incorporated some of the AI-proposed changes into a chip design that has already been sent for manufacturing.

The announcement, published by OpenAI on November 18, 2026, represents a notable step in the use of AI beyond conventional software development. Instead of simply generating code for applications, OpenAI’s models worked on SystemVerilog code that describes how electronic circuits operate. The company says the models identified further optimization opportunities in AMD designs that engineers had already spent years improving.

OpenAI describes the result as roughly a year of additional design progress achieved in less than five months. However, the headline figure requires careful interpretation. The reported 14% is a power, performance and area improvement on AMD GPU intellectual property, not a promise that a finished graphics card will deliver 14% higher frame rates or consume 14% less electricity.

The distinction matters because circuit-level improvements must pass through several stages before their effects can be measured in a commercial product. Even so, AMD’s reported adoption of some proposed changes gives the announcement greater significance than an AI experiment conducted entirely outside a production engineering process.

Key Highlights of OpenAI and AMD’s Chip Design Work

  • OpenAI reports up to 14% improvement in power, performance and area measures on AMD’s flagship GPU intellectual property.
  • The improvements were identified in less than five months against heavily optimized design baselines.
  • AMD has included some AI-proposed circuit changes in a design sent for manufacturing.
  • Other proposed improvements have been added to AMD’s production design files for potential use in future versions.
  • The work involved rewriting SystemVerilog RTL and checking circuit behavior with Synopsys electronic design automation tools.
  • OpenAI also reports single-digit percentage PPA gains across key blocks of an upcoming AMD flagship CPU.

What Does the 14% Improvement Actually Mean?

The central technical measure in OpenAI’s announcement is PPA, which stands for power, performance and area. These three factors are fundamental to processor design because engineers must balance how much energy a circuit uses, how quickly it performs its tasks and how much physical space it occupies on a chip.

Power refers to the electricity consumed by a circuit. Reducing power can help improve energy efficiency and manage the heat generated during operation. This is particularly important in data centres, where thousands of processors may operate continuously and electricity consumption can become a major operating expense.

Performance describes how effectively a circuit carries out its intended work. Depending on the design, improvements may involve completing operations more quickly, increasing throughput or reducing delays in particular parts of a processor.

Area refers to the physical space occupied by the circuit on the silicon die. A smaller circuit can potentially allow manufacturers to fit more functional components into a limited area, improve design flexibility or reduce manufacturing costs. However, the commercial benefit depends on the wider chip architecture and manufacturing process.

These measurements are connected rather than independent. A circuit redesign might reduce area while maintaining performance, or increase performance without raising power consumption significantly. In other cases, improving one measure may require compromises in another.

That is why the reported 14% should not be interpreted as a separate 14% improvement in all three categories. OpenAI describes it as an improvement in PPA relative to heavily optimized baselines, but its announcement does not provide a detailed breakdown showing how much of the result came from power, performance or area.

The company has also not identified the individual GPU blocks responsible for the maximum result or disclosed enough measurement detail to independently reproduce the headline figure. The most accurate conclusion is that OpenAI reports a maximum PPA improvement on selected AMD GPU intellectual property, rather than a measured 14% increase in the performance of an entire graphics processor.

How OpenAI’s AI Models Rewrote AMD’s GPU Circuits

The work focused on SystemVerilog register-transfer-level code, commonly known as RTL. RTL is a hardware description language used by chip designers to describe the operations and data movement that take place inside digital circuits.

Unlike ordinary software, which runs on an existing processor, RTL describes the hardware that will eventually be implemented in silicon. It provides a structured representation of how a circuit handles calculations, controls operations and moves information between different components.

OpenAI says its models examined this code and proposed alternative circuit descriptions. The objective was not simply to shorten the source code or make it easier to read. The models were looking for ways to change the underlying circuit implementation while preserving the required behavior.

For example, a processor block may perform mathematical calculations through a particular arrangement of logic. A different circuit arrangement could potentially perform the same operation using less area, consume less power or complete the operation more efficiently. Finding such alternatives requires understanding both the intended function and the physical consequences of a proposed design.

OpenAI says its models were able to reason about RTL at the level of an expert microarchitect. Microarchitecture concerns the internal organization of a processor, including how its functional units, control logic and data pathways work together.

The models used electronic design automation tools from Synopsys to evaluate their proposed changes. These tools helped verify that a rewritten circuit preserved the behavior of the original block and assess the trade-offs involving power, performance and area.

This verification step is essential. A circuit that appears more efficient but produces incorrect results cannot be used in a production processor. Similarly, an optimization that improves one measurement but creates unacceptable timing or implementation problems may not be suitable for manufacturing.

The process therefore combined AI-generated design alternatives with established engineering tools. OpenAI’s models explored possible improvements, while verification and evaluation tools helped determine whether those changes were technically valid and worth adopting.

Why AMD’s Decision to Use Some Changes Is Important

AI-generated hardware designs are not particularly meaningful to the semiconductor industry unless they can survive the requirements of real chip development. A model can produce an interesting circuit proposal in a controlled environment, but production hardware must satisfy strict requirements for correctness, timing, physical implementation and reliability.

OpenAI says AMD has already included some of the proposed GPU circuit changes in a recent design sent for manufacturing. Other changes have been added to production design files for use in future versions.

This is an important distinction because it moves at least part of the work beyond an experimental demonstration. The proposed changes have entered AMD’s engineering process, and some are part of a design submitted for manufacturing.

However, a design being sent for manufacturing does not mean that finished chips have already been produced, tested and independently benchmarked with the reported improvements. Semiconductor manufacturing and validation involve additional stages, and the announcement does not identify a retail product that buyers can purchase with these specific changes.

OpenAI has also not named the exact AMD GPU architecture or commercial product associated with the work. It has not published a complete list of modified circuit blocks or disclosed the final measured effect on an entire processor.

Consequently, the strongest evidence currently available is the reported adoption of some changes into AMD’s design workflow. The final consumer impact will depend on what AMD manufactures, how the optimized blocks perform in the complete chip and whether the benefits are reflected in shipping products.

OpenAI Also Reports Gains in AMD’s Next-Generation CPU Design

The collaboration extends beyond graphics processors. OpenAI says AMD verified single-digit percentage improvements in PPA across key blocks of its upcoming next-generation flagship central processing unit (CPU), using designs written by OpenAI’s model.

This is a separate result from the up-to-14% GPU figure. The company has not presented the CPU improvements as another 14% gain, and the two figures should not be combined.

CPUs and GPUs perform different but increasingly connected roles in modern computing. A CPU typically handles a broad range of general-purpose operations and coordinates work across a computer system. A GPU is designed to process large numbers of operations in parallel, making it particularly useful for graphics workloads and many artificial intelligence calculations.

Improving the internal design of either processor type can have implications for computing efficiency. For data-centre operators, better power and performance characteristics can influence the amount of computing capacity available within limits on electricity, cooling and physical infrastructure.

OpenAI has not identified the specific AMD CPU product or published a detailed breakdown of the verified gains. The announcement therefore establishes that the collaboration has produced results in both GPU and CPU design work, but it does not yet provide enough information to estimate the impact on future servers or consumer computers.

OpenAI’s Jalapeño Chip Shows a Different Stage of AI-Assisted Design

The AMD collaboration is part of a broader effort by OpenAI to use its models in semiconductor engineering. The company also discussed improvements to Jalapeño, its first custom inference chip, in the same research announcement.

Inference is the process of using a trained AI model to generate an output, such as answering a question, producing an image or completing another task. The hardware used for inference must deliver the required computing capacity while managing power consumption, performance and cost.

OpenAI says it began using GPT-4o to help design Jalapeño in early 2025. It reports that the model-proposed optimizations were silicon-proven during the chip’s bring-up process, the stage when engineers power up newly manufactured silicon and test whether it operates as intended.

The company describes improvements to Jalapeño’s compute blocks, including its vector and matrix units. These are important components for performing different types of mathematical operations used in computing and AI workloads.

The Jalapeño results provide a useful comparison with the AMD announcement. OpenAI says the Jalapeño optimizations were verified during silicon bring-up, while the AMD GPU announcement describes some changes entering a design sent for manufacturing and others being added to production design files.

These are different stages of evidence. Silicon bring-up provides information about actual manufactured hardware, whereas a manufacturing submission indicates that a design has advanced through an important part of the development process but does not by itself establish final product performance.

OpenAI Tests Whether Its Models Can Design Processor Architectures

OpenAI also evaluated successive generations of its models on two open-source 32-bit RISC-V processor designs: PicoRV32 and the Ibex core used in OpenTitan. These tests were intended to measure how model capabilities have progressed on end-to-end chip design tasks.

RISC-V is an open instruction set architecture that allows organizations and developers to design processors without relying on a single proprietary instruction set. PicoRV32 is a compact processor core, while Ibex is used as the embedded processor within the OpenTitan security-focused hardware project.

For the evaluations, OpenAI provided models with the original hardware descriptions and asked them to improve throughput. Throughput measures how much work a processor can complete over a period of time, such as the number of instructions it executes per second.

Improving throughput is not always as simple as reducing the time needed for each individual operation. A design might complete an instruction in fewer clock cycles but also make each clock cycle longer. A useful evaluation must consider both effects rather than focusing on one measurement in isolation.

OpenAI checked the resulting designs through simulation and applied an area limit to prevent models from achieving better results simply by creating excessively large circuits. Designs that failed the required checks or exceeded the specified area threshold received no score.

However, OpenAI also acknowledged limitations in the evaluation. Starting from written requirements gave the models room to omit features present in the original processor designs, including performance counters and additional instructions. Such features may not contribute much to throughput on the particular test workloads, but they can still matter in a complete production processor.

The company noted that requiring those features, applying tighter circuit-area limits and checking timing after physical synthesis would make the evaluation more representative of production chip design.

This qualification is important because benchmark performance does not automatically translate into manufacturing-ready hardware. Real processors must satisfy a much wider set of functional and physical requirements than a limited test can capture.

Why the Partnership With Synopsys Matters

OpenAI is also working with Synopsys to connect its model capabilities with the electronic design automation tools used by semiconductor engineers. The companies are developing GPT-Synopsys, a model intended to use Synopsys design and validation tools in ways that support chip development.

Electronic design automation software plays a central role in modern semiconductor engineering. Designing a complex processor involves managing enormous numbers of interconnected components, checking whether circuits behave correctly and evaluating whether a design can meet its physical implementation requirements.

AI models could potentially help engineers explore more design alternatives, automate repetitive tasks and identify optimization opportunities that might otherwise require substantial manual investigation. But the tools must work within established verification processes because even a small design error can have serious consequences in manufactured hardware.

The Synopsys collaboration reflects an approach in which AI is integrated into existing engineering workflows rather than treated as a replacement for the entire chip-design process. Human engineers, specialized software and formal verification remain important parts of the process.

For the semiconductor industry, this combination could become increasingly valuable as processor designs grow more complex and the cost of developing advanced chips continues to demand careful engineering decisions.

What the AMD GPU Improvements Could Mean for AI Data Centres

The potential benefits extend beyond gaming graphics cards. AMD’s flagship GPU intellectual property is relevant to the broader high-performance computing market, where GPUs are used for demanding workloads, including artificial intelligence training and inference.

In large data centres, power efficiency is a major consideration. Operators must supply electricity to processors, networking equipment, storage systems and cooling infrastructure. Improving the efficiency of individual computing components can help increase the amount of useful work performed within a fixed power budget.

Area improvements may also matter to chip manufacturers because silicon space is a limited resource. Depending on the design and manufacturing process, more efficient circuit blocks can create opportunities to adjust the chip’s architecture or allocate space to other functions.

Performance improvements could help processors complete certain tasks more efficiently. However, the overall effect depends on the workload and the rest of the system. A faster circuit block does not necessarily make every application faster if another component becomes the main performance limitation.

For OpenAI, the work also has strategic relevance. The company depends on large amounts of computing capacity to develop and operate AI systems. Improving the hardware used for those workloads could support better efficiency across its infrastructure over time.

OpenAI and AMD already have a wider commercial relationship. In October 2025, the companies announced a multiyear agreement involving the deployment of up to six gigawatts of AMD GPUs for OpenAI’s AI infrastructure, with the initial deployment planned around AMD Instinct MI450-series products. The circuit-design collaboration adds another dimension to that relationship by focusing on improving the underlying hardware design process.

What the Announcement Does Not Yet Prove

Despite the promising engineering results, several important questions remain unanswered. OpenAI has not disclosed the exact AMD GPU product associated with the 14% result, the individual blocks that achieved the maximum improvement or the detailed calculation behind the headline figure.

It has also not published a complete breakdown of the power, performance and area changes. Without those figures, it is difficult to determine whether the most significant benefit comes from lower power consumption, improved circuit speed, reduced area or a combination of trade-offs.

Another important question concerns finished silicon. A design sent for manufacturing is a meaningful milestone, but independent testing of completed chips would provide stronger evidence of the practical benefits. Such testing could establish how the modified circuits behave under real operating conditions and whether the expected improvements remain measurable in the final product.

There is also no confirmed retail graphics card or consumer processor that buyers can associate with the reported gains. AMD has not identified a specific product launch, price change or performance benchmark tied to the AI-proposed GPU modifications in OpenAI’s announcement.

For these reasons, consumers should not interpret the announcement as evidence that upcoming Radeon graphics cards will automatically be 14% faster or cheaper. The result is primarily a development in semiconductor engineering, and its commercial consequences will become clearer only when AMD provides product-level information.

What Happens Next in AI-Assisted Chip Design?

The next stage will involve determining how consistently AI models can deliver improvements across different processor architectures and increasingly demanding engineering tasks. OpenAI says it expects further PPA improvements on top of the optimized GPU and CPU baselines.

That expectation is forward-looking rather than a confirmed result. More work will be needed to establish how broadly the approach can be applied, how much engineering review it requires and whether the improvements remain reliable across different designs and manufacturing conditions.

The partnership with Synopsys could also help expand the use of AI in chip development by connecting model capabilities with tools that engineers already use. If these systems can reliably propose valid optimizations and reduce the time needed to evaluate them, they could become useful assistants throughout the semiconductor design process.

Still, AI-assisted chip design will not eliminate the need for experienced engineers. Semiconductor development requires decisions about architecture, physical implementation, manufacturing constraints, verification and product requirements. Models can help explore possible solutions, but those solutions must still pass rigorous technical checks before they can become part of a finished processor.

Conclusion: A Significant Engineering Milestone, Not a 14% Faster GPU Promise

OpenAI’s reported 14% PPA improvement on AMD’s flagship GPU intellectual property is an important development in the use of AI for semiconductor engineering. The achievement is notable because the models worked on designs that had already undergone extensive optimization, and AMD has reportedly incorporated some proposed changes into a design sent for manufacturing.

The collaboration also extends to AMD’s next-generation flagship CPU, where OpenAI reports single-digit percentage PPA gains across key blocks that AMD verified. Together with the Jalapeño work and the planned collaboration with Synopsys, these results show how AI models are being tested in increasingly practical hardware-design environments.

However, the reported GPU figure should remain in its proper context. It does not establish a 14% increase in the speed of a complete graphics card, a guaranteed reduction in electricity consumption or an immediate change in retail prices. The exact product impact remains unknown until the relevant hardware is manufactured, tested and identified.

The larger significance is that AI models are beginning to contribute to the engineering of the processors that power modern computing. If the approach continues to produce reliable improvements, it could change how chipmakers search for efficiency gains and develop future generations of CPUs and GPUs.

FAQs

  • What 14% improvement did OpenAI report in AMD GPU design?
  • Does the 14% improvement mean AMD graphics cards will be 14% faster?
  • Has AMD used OpenAI's AI-generated chip design improvements?
  • How did OpenAI use AI to improve AMD GPU circuits?
  • What does PPA mean in semiconductor chip design?
  • Did OpenAI also improve AMD CPU designs?
  • What is OpenAI's Jalapeño chip?
  • When will consumers see the benefits of OpenAI's AMD GPU design improvements?

For breaking news and live news updates, like us on Facebook or follow us on Twitter and Instagram. Read more on Latest Technology on thefoxdaily.com.

COMMENTS 0