OpenAI said it improved AMD's flagship GPU design by up to 14 percent using an AI model. [Photo: Shutterstock]

OpenAI said it deployed its artificial intelligence (AI) model in semiconductor design and improved AMD's flagship graphics processing unit (GPU) design by up to 14 percent. It marks AI beginning to directly participate in designing chips that will train and run next-generation AI models.

Cryptocurrency outlet Cryptopolitan reported on Nov. 19 that OpenAI said in research released on Nov. 18 that applying its model to AMD's highly optimised flagship GPU intellectual property (IP) design improved power, performance and area (PPA) metrics by up to 14 percent. It did not disclose the gains by PPA category or which AMD products were involved.

OpenAI said some parts of the AMD chip design rewritten by the model were already reflected in designs submitted for manufacturing. The rest were included in production design files so they can be applied to future products. AMD is also reported to have confirmed single-digit percent PPA improvements in major blocks of its next-generation flagship central processing unit (CPU) using designs produced by the model.

OpenAI stressed it achieved about a year of design progress in less than 5 months in its work related to AMD. The model rewrote SystemVerilog RTL code describing chip circuits and used Synopsys verification tools to confirm the functional integrity of circuit blocks after changes.

The approach was also applied to its in-house inference chip, Jalapeno. OpenAI said it has designed Jalapeno using GPT-4o since early 2025. Changes proposed by the model reduced standard cell area by about 13 percent for the vector unit and about 10 percent for the matrix unit, and the improvements were confirmed in an actual silicon validation process.

To evaluate the model's semiconductor design capabilities, OpenAI tested five generations of models on PicoRV32 and Ibex, open-source RISC-V processor cores. Across a total of 9,881 evaluations, GPT-6.2 Astra recorded the highest average score on both cores. It said the results were based on simulation and do not mean performance improvements in mass-produced chips.

The next step is semiconductor design automation. OpenAI is developing a specialised model called GPT-Synopsys in cooperation with Synopsys that directly uses electronic design automation (EDA) tools. That would expand AI's scope from circuit design to operating specialised design tools.

Risks are also being raised over AI becoming involved in next-generation AI development. OpenAI said its preparedness framework manages AI self-improvement as one of the key capabilities that could pose a major risk. It said rapid improvements in AI performance could create new difficulties for humans in controlling systems.

OpenAI said it would halt development if its model reaches a “major threshold” in self-improvement until appropriate safeguards are in place. However, researcher Charbel Raphael (샤르벨 라파엘) criticised in a May 2026 analysis that the standard is triggered too late and relies on self-evaluation. He also pointed out that OpenAI has not clearly defined specific criteria for mitigating major risks.

Keyword

#OpenAI #AMD #Synopsys #GPT-4o #RISC-V
Copyright © DigitalToday. All rights reserved. Unauthorized reproduction and redistribution are prohibited.