🔍 Read the full analysis: The Context Behind OpenAI’s Up To 14% AI Improvement For AMD Designs on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI said on November 18 that its models rewrote parts of AMD chip designs, reporting up to 14% gains in power, performance and area for a flagship GPU against heavily optimised baselines. AMD is reported to have verified single-digit gains in key blocks of a next-generation CPU, but the results and measurement details have not been independently published. Some changes are said to be in a design sent for manufacturing.
OpenAI said November 18 that its models rewrote parts of AMD chip designs, producing up to 14% improvements in power, performance and area (PPA) on a flagship GPU compared with heavily optimised baselines. OpenAI also said some model-generated changes were included in an AMD design sent for manufacturing, while AMD had verified single-digit PPA gains in key blocks of a forthcoming CPU. The figures have not been independently audited in public, and the companies have not provided enough detail to establish how broadly they apply.
OpenAI announced the work in a post titled “Advancing chip design with OpenAI models.” The reported results cover several distinct projects, rather than one overall improvement across AMD products. For the flagship GPU, OpenAI described a best-case gain of up to 14% in PPA versus what it called heavily optimised baselines. It did not identify the specific GPU blocks, give an average gain, or state what proportion of the chip those blocks represent.
For a next-generation AMD CPU, OpenAI said AMD had verified single-digit percentage gains in performance, power and area in key blocks. The reports do not include figures from an AMD technical publication or a detailed account of the verification. OpenAI also said some generated changes were included in a recent design sent for manufacturing, with other changes added to design files for future products. The company did not specify which chip or changes were in the manufacturing submission.
OpenAI described a separate result on its own custom inference chip, called Jalapeño. It said model-proposed changes reduced standard-cell area by about 13% in one unit and 10% in another, with those results validated after the first chips came back. Accounts of the first unit differ: ANI described it as a vector unit, while Inside AI called it a warp engine. These Jalapeño figures are not AMD results and should not be conflated with the GPU’s up-to-14% headline.
AI Enters the Chip Design Workflow
The report points to AI being used not only to develop software and models, but also to alter the hardware designs on which future AI systems run. If the changes perform as reported after manufacturing, reductions in chip area or improvements in power and speed could affect how efficiently a processor uses silicon and energy. The practical value depends on which blocks changed, how large they are relative to the whole chip, and whether the gains survive later design and manufacturing stages.
The reported process also assigns different roles to models and established engineering tools. OpenAI said its models proposed alternative register-transfer-level (RTL) code, while Synopsys design tools checked that rewrites preserved behaviour and assessed PPA. That distinction matters because a more compact or faster circuit is not useful if it produces incorrect results. Conventional verification is therefore part of the claim, rather than evidence that the model alone established correctness.
There is also a commercial reason to read the figures carefully. OpenAI has a financial relationship with AMD: according to the source material, an agreement announced in October 2025 gave OpenAI a warrant for up to 160 million AMD shares, vesting as AMD chips are deployed. That relationship does not disprove the engineering claims, but it makes clear attribution and independent technical evidence particularly relevant. The public information currently comes largely from OpenAI’s account.
Popular right now

TCL 75-Inch QM8L Series 4K QLED Smart Google TV

KitchenAid 5.5 Qt Bowl-Lift Stand Mixer – Porcelain White (KSM55SXXX)

iRobot Roomba Plus 405 Combo Robot Vacuum and Mop with AutoWash Dock
As an affiliate, we earn on qualifying purchases.
From RTL Rewrites to Silicon
The work described takes place at the RTL stage of chip design. Engineers use RTL code, commonly written in SystemVerilog, to specify how circuit blocks behave over clock cycles. Tools then translate that description into logic, followed by later steps that place and connect components on a chip. OpenAI’s reported contribution was to have models propose alternative RTL implementations of existing functions.
According to OpenAI, as reported by ANI and other outlets, Synopsys tools checked whether proposed changes retained the original behaviour and assessed their power, performance and area. The source accounts describe possible strategies such as changing logic, dividing calculations differently across pipeline stages, sharing hardware or reducing intermediate storage. Inside AI also described Jalapeño examples: a multiplier redesign reportedly used about 15% less cell area in synthesis while reducing latency from two clock cycles to one; an adder redesign reportedly cut cell area by about 15% at the same latency. Those are examples from OpenAI’s chip, not independently reported AMD measurements.
OpenAI also said it tested five models on two open-source processor cores, PicoRV32 and Ibex, across 9,881 scored runs. The source material reports that GPT-6.2 Astra received the highest score. This benchmark is separate from the AMD product claims: success on small open-source cores does not, by itself, establish the scale or manufacturing impact of improvements to a commercial GPU or CPU.
The Missing Measurement Details
The public accounts do not identify the specific AMD GPU blocks behind the up-to-14% result, their share of the chip, or whether the figure represents a synthesis-stage estimate or measurements after placement and wiring. No typical or average result is given. Without those details, readers cannot tell whether the best case is representative of the design or limited to a small section.
It is also not clear which model-generated changes went into the design sent for manufacturing, what stage that design had reached, or whether the reported gains were confirmed on manufactured AMD silicon. OpenAI says the Synopsys tools checked functional equivalence and assessed PPA, but the underlying test results and methodology have not been made public in the source material. AMD has not published its own figures, and no independent audit of the results is described.
OpenAI’s financial relationship with AMD is disclosed in the source material, but the exact commercial and technical arrangements connected to this design work are not detailed there. The available information relies principally on OpenAI’s announcement and news reports of it, so the distinction between the company’s claims and independently verifiable results remains important.
Evidence Needed From AMD and OpenAI
The next useful evidence would be a technical account from AMD or OpenAI identifying the affected design blocks, the measurement stage, the comparison baselines and the range of results rather than only the maximum. AMD could also clarify which changes were included in the design sent for manufacturing and whether later tests confirmed the gains. No such publication or specific release date is provided in the source material.
As the reported designs progress, manufacturing and validation results could show whether RTL-level improvements persist in physical chips. Until those details are available, the figures should be treated as company-reported design results, not proof of a whole-chip improvement or a confirmed gain in deployed products.
Key Questions
What does OpenAI’s up-to-14% figure refer to?
It refers to OpenAI’s reported best-case improvement in power, performance and area for parts of an AMD flagship GPU design, compared with heavily optimised baselines. OpenAI has not published an average or identified the blocks behind the figure in the source material.
Did AMD confirm the reported results?
OpenAI said AMD had verified single-digit PPA gains in key blocks of a next-generation CPU. AMD has not published its own detailed measurements in the supplied source material. The separate up-to-14% GPU figure is also presented through OpenAI’s account.
Were the new designs already manufactured?
OpenAI said some model-generated changes were included in a design AMD had sent for manufacturing. The announcement as described does not identify the design, say which changes it contains, or confirm that manufactured chips have demonstrated the claimed gains.
How did the models contribute to chip design?
OpenAI said its models proposed rewrites of SystemVerilog RTL, the code describing circuit behaviour. Synopsys tools, rather than the model alone, checked that changes preserved the original behaviour and assessed their PPA effects.
Has an independent audit been published?
No independent audit of the AMD chip-design results is described in the source material. The public claims rely mainly on OpenAI’s announcement as reported by news outlets, with detailed methods and results still unavailable.
Source: ThorstenMeyerAI.com
