🔍 Read the full analysis: What OpenAI Measured When Its Models Improved AMD Chip Designs By Up To 14% on ThorstenMeyerAI.com
Get movie nights 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 its models rewrote parts of AMD chip designs and produced up to 14% gains in power, performance and area against heavily optimized baselines. The company also reported smaller area reductions on its own Jalapeño inference chip, while AMD has not published its own figures and no independent audit of the results is available.
OpenAI said on November 18 that its models rewrote parts of AMD chip designs and produced up to 14% improvement in power, performance and area, or PPA, against heavily optimized baselines. The announcement matters because it describes AI being used to refine hardware for future computing, but the headline figure is a best-case result reported by OpenAI, with limited public detail and no independent audit.
OpenAI published the results in a post titled “Advancing chip design with OpenAI models.” The company said some model-generated changes were included in a recent AMD design sent for manufacturing, while other changes were added to production design files for future versions. The public reports do not identify the affected blocks or say what proportion of the chip they represent.
OpenAI also said AMD verified single-digit PPA gains in key blocks of its next flagship CPU. AMD chief executive Lisa Su praised the work on X, according to Stocktwits, but AMD has not published its own figures. OpenAI president Greg Brockman described the work as “roughly an additional year of progress” achieved in less than five months, a characterization of the pace that remains OpenAI’s own.
The reported work focused on register-transfer-level (RTL) code, which describes how chip circuits behave. Models proposed alternative RTL implementations; Synopsys design tools then checked whether the changes preserved the original behavior and assessed their PPA, according to OpenAI as reported by ANI and other outlets. The models did not certify their own rewrites.
OpenAI gave separate results for its own Jalapeño inference chip: it said model-proposed changes cut standard-cell area by about 13% in one unit and 10% in another, with the changes validated after the first chips came back. Reports differ on the first unit’s name, calling it either a vector unit or a warp engine. Inside AI also described a multiplier redesign that 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 latency.
AI Moves Into Chip Design
If the reported gains hold through manufacturing and operation, AI-assisted RTL work could help chip teams improve designs without relying solely on engineers to handcraft each circuit change. Even modest improvements can matter in hardware, where power, performance and area compete: lower area may reduce cost, while better performance or power efficiency can affect the capabilities and running costs of systems built around a chip.
The potential is especially relevant to AI infrastructure. More efficient processors could change how much computing capacity can be delivered within limits on power, cooling and manufacturing supply. But the announcement does not establish that every AMD product will improve by 14%, or that the reported best case translates into a comparable gain in a finished chip. The real-world effect depends on which blocks changed, how large they are, and whether improvements survive later design and manufacturing steps.
The results also illustrate a possible feedback loop: AI models help design the hardware used to run AI systems. That could shorten some engineering tasks, but chip development still depends on conventional verification, physical design and manufacturing. OpenAI’s description is evidence of reported experiments, not a public demonstration that the full process can be automated.
Popular right now

Bose QuietComfort Wireless Noise Cancelling Headphones – Twilight Blue Limited Edition

Philips 5500 Series Fully Automatic Espresso Machine with LatteGo Milk System

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black
As an affiliate, we earn on qualifying purchases.
From RTL Rewrites to Silicon
RTL is an early description of a chip’s logic. Engineers specify what circuit blocks should do; synthesis tools translate that code into logic gates, and later stages place and connect those gates on silicon. Changes that look equivalent at the RTL stage must still meet constraints through later steps. A design can be faster in one measure but worse in another, or fail to preserve the intended behavior.
OpenAI said Synopsys tools checked that the proposed rewrites preserved the original design’s behavior and measured their PPA effects. That verification step matters because a subtle hardware bug can require a costly redesign or manufacturing run. OpenAI’s account, as summarized by ANI, The Decoder and Inside AI, says the AMD comparison used baselines that AMD engineers had already optimized extensively. That makes the comparison more meaningful than one against an unoptimized design, but the public material does not provide enough detail to reproduce it.
OpenAI also reported a separate benchmark involving two open-source processor cores, PicoRV32 and Ibex. It said five models completed 9,881 scored runs, with GPT-6.2 Astra scoring highest. That benchmark is distinct from the AMD chip results and should not be treated as independent confirmation of the 14% claim. The source material says the model was released the day before the announcement, but gives no comparable AMD-specific test data from that benchmark.
Details Behind the 14% Figure
The public account does not name the AMD GPU blocks that produced the best result, show how much of the chip they represent, or give an average across the tested changes. “Up to 14%” is a maximum, not a typical result. The reports also do not specify whether the GPU figure was measured after synthesis or after placement and wiring, later stages that can change the final trade-offs.
No independent audit of the AMD or Jalapeño results has been published. The key figures come from OpenAI’s announcement as relayed by news outlets; AMD’s public response, as described in the source material, includes praise from Su but no detailed results. The two companies also have a financial relationship: since October 2025, OpenAI has held a warrant for up to 160 million AMD shares that vests as AMD chips are deployed. That relationship does not disprove the engineering claims, but it is relevant when weighing company-reported results.
It is also unclear how many model-generated proposals were tested, how many were rejected, and whether the stated improvements were measured under identical constraints across designs. Public information does not establish how the changes perform in volume-produced chips or what effect they have on overall system performance.
Evidence Needed From AMD
The next useful evidence would be more detailed figures from AMD or OpenAI: the affected design blocks, the baseline and measurement stage, the range of results rather than only the best case, and whether gains persisted through physical design and manufacturing. AMD’s own technical disclosures could clarify what “single-digit” gains mean for the CPU and how much of its design was changed.
For the GPU changes reportedly sent for manufacturing, the announcement does not give a release schedule or identify a product that will contain them. Further validation may emerge as AMD discusses future chips or publishes engineering material, but the source material does not state when that might happen. Until then, the 14% figure remains a company-reported best case rather than a publicly reproducible measure of a finished chip’s performance.
Key Questions
What does the 14% improvement refer to?
OpenAI said its models achieved up to 14% improvement in power, performance and area on parts of an AMD flagship GPU design, compared with heavily optimized baselines. The company did not say which blocks produced the best result or provide an average.
Did AMD confirm the 14% figure?
The available reporting attributes the GPU figure to OpenAI. OpenAI separately said AMD verified single-digit gains in key blocks of a next flagship CPU. AMD has not published its own detailed figures in the source material.
How did the models change the chip designs?
The models proposed changes to RTL, the code that describes circuit behavior. According to OpenAI, Synopsys tools checked functional equivalence and assessed the power, performance and area effects of those rewrites.
Have the changes reached manufactured chips?
OpenAI said some changes were included in an AMD design sent for manufacturing. It did not identify the product, say when it will be available, or report measured performance from a commercial chip.
Are the results independently verified?
No independent audit of the results has been published in the source material. The reported figures come from OpenAI’s announcement, with AMD’s own public technical data still limited.
Source: ThorstenMeyerAI.com
