🔍 Read the full analysis: OpenAI And AMD Chip Design: Understanding The Up To 14% Result on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI said on November 18 that its models produced up to 14% improvements in power, performance and area on AMD’s flagship GPU design, with some changes included in a design sent for manufacturing. The figure is a best-case result reported by OpenAI; the companies have not published block-level measurements or an independent audit.
OpenAI said November 18 that its models rewrote parts of designs for AMD’s flagship graphics processor, producing up to 14% gains in power, performance and area (PPA), according to the company’s account of the work. OpenAI said some model-generated changes are in an AMD design sent for manufacturing, but AMD has not published its own figures, and no independent audit of the results has been reported.
The reported gains came from models proposing changes to SystemVerilog register-transfer-level (RTL) code, which describes how a chip’s circuit blocks behave. OpenAI said conventional Synopsys design tools checked that the revised code preserved the original behavior and measured its PPA trade-offs. The model proposed the changes; the reported functional and engineering checks were performed with design tools.
OpenAI described the AMD GPU result as “up to” 14% against heavily optimized baselines. It said some changes were included in a recent design sent for manufacturing, while others went into production design files for future versions. The company also said AMD had verified single-digit percentage gains in performance, power and area in key blocks of its next flagship CPU. AMD has not released detailed measurements of its own, according to the source material.
OpenAI reported separate results on its own Jalapeño inference chip: model-proposed changes reduced standard-cell area by about 13% in one unit and 10% in another, with the results validated after the first chips came back, it said. Reports differ on the name of the first unit, calling it a warp engine or a vector unit. OpenAI also described tests involving five models and two open-source RISC-V processor cores, PicoRV32 and Ibex, across 9,881 scored runs. These results are distinct from the AMD GPU figure.
AI Models Enter Chip Design
If the reported changes hold up in manufactured products, they could help chip teams improve performance, power use or silicon area without changing what a circuit is meant to do. Those trade-offs affect the cost and capability of processors used to run AI systems. The work also points to a feedback loop: AI models may help design the hardware used to train or operate later AI models.
The practical reach of the result is not yet clear. The 14% is a best-case figure, not a disclosed average across the GPU, and the companies have not identified which blocks produced it or how much of the chip they represent. A gain measured early in the design process may not translate unchanged to a finished, manufactured processor.
The companies also have a financial relationship relevant to readers’ evaluation of the claims. Since October 2025, OpenAI has held a warrant for up to 160 million AMD shares that vests as AMD chips are deployed, according to the source material. That tie does not establish whether the engineering results are accurate, but it makes independent measurement and clear disclosure especially relevant.
Popular right now

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

DREAME D30 Ultra CE Robot Vacuum and Mop, 25,000Pa Suction

LG OLED evo C4 55-Inch 4K Smart OLED TV (OLED55C4PUA) with 144Hz Refresh Rate, HDMI Cable, 24-Month Protection Plan, and Streaming Bundle
As an affiliate, we earn on qualifying purchases.
From RTL Code to Silicon
RTL is an early stage of chip design. Engineers use it to describe a circuit’s operations cycle by cycle; synthesis tools then translate that description into logic gates, followed by steps that place and connect components on a chip. OpenAI’s reported work addresses the RTL stage, where alternative logic or pipeline arrangements may implement the same function with different trade-offs.
The source material describes conventional tools checking that model-generated rewrites preserved behavior. That distinction matters because a design that appears faster or smaller but changes the circuit’s function cannot be used safely. OpenAI said Synopsys tools performed the verification and assessed PPA; the models did not establish correctness on their own.
OpenAI president Greg Brockman characterized the pace of the AMD work as “roughly an additional year of progress” achieved in less than five months. That is OpenAI’s description of the outcome, not an independently verified estimate of engineering time saved. The reported comparisons were against AMD designs that had already been heavily optimized, according to coverage of OpenAI’s post.
The Missing Measurement Details
The reports do not identify the AMD GPU blocks that produced the 14% best-case result, their share of the full chip, or whether the measurement came after synthesis or after placement and wiring. They also do not provide an average result across the design or enough detail to reproduce the comparison. The baseline is described as heavily optimized, but the underlying measurements have not been published.
OpenAI says some changes are in a design sent for manufacturing, but the available source material does not specify which changes, the product’s manufacturing timeline, or whether the claimed gains have been confirmed in production silicon. AMD’s reported single-digit CPU gains and OpenAI’s Jalapeño results have the same limitation: detailed public data and independent audits are not available. OpenAI’s financial interest in AMD is also relevant disclosure, not evidence for or against the technical claims.
Manufacturing Results and Public Data
The next meaningful test will be whether the design sent for manufacturing produces silicon that retains the reported gains. Further detail from AMD or OpenAI could clarify which circuit blocks changed, how the measurements were made, and whether results reflect the complete chip or selected blocks.
For now, readers should treat the figures as company-reported engineering results, not independently audited product benchmarks. The source material does not give a date for additional disclosures or a specific release milestone for the AMD design.
Key Questions
What does the up-to-14% figure measure?
OpenAI says it measures improvements in power, performance and area on parts of AMD’s flagship GPU design, against heavily optimized baselines. It has not disclosed the average result or the blocks involved.
Did AMD confirm the 14% result?
The source material attributes the GPU figure to OpenAI. OpenAI also said AMD verified single-digit PPA gains in key blocks of a next flagship CPU, but AMD has not published its own detailed figures.
How did the models change the chip designs?
They proposed rewrites to SystemVerilog RTL code. OpenAI said Synopsys design tools checked that the revised code preserved the original behavior and measured engineering trade-offs.
Are the reported gains confirmed in manufactured chips?
OpenAI said some changes were included in an AMD design sent for manufacturing. The available source material does not say whether finished AMD silicon has confirmed the claimed gains.
What remains unknown about the result?
The companies have not disclosed the affected GPU blocks, their size relative to the full chip, an average across the design, or enough measurement detail for an independent audit. The 14% figure is a reported maximum, not a stated typical result.
Source: ThorstenMeyerAI.com
