🔍 Read the full analysis: OpenAI Says Its Models Helped AMD Chip Designs. How Was That Judged? on ThorstenMeyerAI.com
Get the latest gadgets 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 against heavily optimized baselines. Conventional Synopsys tools checked the rewrites, according to OpenAI, but the companies have not published block-level results or an independent audit. AMD has a financial relationship with OpenAI tied to future chip deployments.
OpenAI said November 18 that its AI models rewrote parts of AMD chip designs, producing up to 14% gains in power, performance and area, or PPA, against what it described as heavily optimized baselines. The company said Synopsys design tools checked whether the changes preserved the original circuits’ behavior, but it has not published enough detail for outsiders to verify the headline figure or judge how representative it is.
OpenAI described results across several projects, not one single test. For an AMD flagship GPU, it reported a best-case improvement of up to 14% in PPA. OpenAI said some model-generated changes were included in a recent design AMD sent for manufacturing, while others went into design files for future versions. The public reports do not identify the affected blocks or state what fraction of the GPU they represent.
For AMD’s next flagship CPU, OpenAI said AMD had verified single-digit percentage gains in PPA in key blocks. AMD has not released its own figures, according to the source material. OpenAI also reported results on its own Jalapeño inference chip: model-proposed changes reduced standard-cell area by about 13% in one unit and 10% in another. The reports differ on the first unit’s name, calling it a warp engine or a vector unit.
The work focused on register-transfer-level (RTL) code, which describes how chip circuits behave. OpenAI said Synopsys tools checked that revised code retained the original behavior and measured its PPA effects. That verification was performed by conventional engineering software, rather than established solely by the AI model. OpenAI also described a test on two open-source RISC-V processor cores, PicoRV32 and Ibex, involving 9,881 scored runs across five models. The supplied material does not provide enough benchmark detail to compare those results with the AMD figures.
AI Moves Into Chip Design
If the reported gains hold through manufacturing, AI-assisted RTL work could help chip teams improve designs without relying only on additional manual redesign. Power, performance and area constrain how much computing a chip can deliver and how efficiently it can run. OpenAI’s account also describes the use of AI tools in designing hardware intended to train and run future AI systems.
The reported evidence is limited. “Up to 14%” is a best-case result, not a stated average across the GPU, and OpenAI has not disclosed the affected blocks or the measurement stage. The company said some changes reached a design sent for manufacturing, but it has not reported what those changes will deliver in a finished product.
OpenAI’s financial relationship with AMD is also relevant to the claims. Since October 2025, OpenAI has held a warrant for up to 160 million AMD shares, vesting as AMD chips are deployed, according to the source material. The relationship does not establish whether the engineering results are accurate; independent scrutiny and documented methods would help readers assess them.
Products Worth Considering

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue

Logitech MX Creative Console with 9 Customizable Keys, Graphite

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From RTL Rewrites to Silicon
Chip design moves through several stages. At the RTL stage, engineers use code to specify a circuit’s operations cycle by cycle. Synthesis tools convert that description into logic gates; later processes place and connect those gates on silicon. A change that looks beneficial in an early step may not preserve its advantage after later design stages.
OpenAI’s reported method had models propose alternative RTL implementations, then used Synopsys tools to check behavior and assess trade-offs. The source material describes examples on Jalapeño, including a multiplier redesign that reportedly used about 15% less cell area in synthesis while reducing latency from two clock cycles to one, and an adder redesign that cut cell area by about 15% at the same two-cycle latency. These are company-reported examples, not independently audited results.
OpenAI published a post titled “Advancing chip design with OpenAI models” on November 18. The source article says its automated request to that post was refused; its account of the post relies on reporting by ANI, The Decoder and Inside AI, which it says agree on the key figures. OpenAI president Greg Brockman summarized the AMD work on X as “roughly an additional year of progress on @AMD’s chip achieved in less than 5 months using our models,” as quoted by Stocktwits. That is a characterization of the timeline, not a separately stated engineering benchmark.
Key Measurement Details Missing
The public account does not name the GPU blocks that produced the reported 14% maximum, their share of the chip, or whether the measurement came after synthesis or after placement and wiring. It also gives no average result across tested changes and does not spell out the precise PPA trade-offs behind the combined figure. Those omissions make it difficult to assess how much of the chip could benefit or whether the gain will survive later design stages.
The source material reports no independent audit of the AMD or Jalapeño results. AMD has not published its own detailed measurements, and the supplied account does not include a full benchmark methodology, underlying data or a clear explanation of the variation in the name of one Jalapeño unit. It is also not yet clear which model produced each AMD change, how many proposed rewrites were rejected, or what performance the manufactured design will achieve.
OpenAI’s financial relationship with AMD is disclosed in the source material, but the public information does not explain how that relationship affected the work or the reporting of results. These gaps do not establish that the claims are inaccurate; they limit what readers can independently conclude from them.
Manufacturing Results and Disclosure
The next useful evidence would be results from the AMD design now reported to be in manufacturing, including measurements from finished silicon and details about which model-generated changes were retained. AMD or OpenAI could also publish block-level results, the measurement stage, the comparison baselines and enough methodology for outside engineers to assess the claims.
For now, the public development is that OpenAI has described AI-generated RTL changes and validation using Synopsys tools, while saying AMD incorporated some changes in a design sent for manufacturing. Real-world chip performance remains unreported in the supplied material. Any assessment of the size or generality of the gains depends on further evidence.
Key Questions
What did OpenAI say its models changed?
OpenAI said its models rewrote parts of AMD GPU and CPU designs, as well as parts of its own Jalapeño inference chip, by proposing alternative RTL code for chip circuits.
What does “up to 14%” mean here?
It is OpenAI’s reported best-case PPA improvement for an AMD flagship GPU compared with heavily optimized baselines. The company has not disclosed an average, the specific blocks, or their share of the chip.
Who checked that the rewritten circuits still worked?
OpenAI said Synopsys design tools verified that changes preserved the original behavior and assessed power, performance and area. The account does not describe an independent audit of the overall results.
Has AMD confirmed the reported gains?
OpenAI said AMD verified single-digit gains in key blocks of its next flagship CPU. The source material says AMD had not published its own detailed figures; it does not provide a separate AMD measurement report.
What remains unknown?
Public details do not identify the relevant GPU blocks, show the full methodology or establish whether the reported gains persist through manufacturing. Finished-chip results and independent verification have not been provided in the source material.
Source: ThorstenMeyerAI.com
