On November 18, 2026, OpenAI announced that AMD had adopted some of its AI-proposed circuit changes in a chip design that AMD sent to manufacturing. The target was a design asset for AMD's flagship GPU. Compared with a design that had already been optimized, the changes delivered improvements of up to 14% on the PPA metrics, which evaluate power consumption, performance and area. OpenAI also said it confirmed changes such as shrinking arithmetic circuits on its own inference chip, "Jalapeño," when the actual chip was brought up. In every case, what the AI modified was RTL (register-transfer level) design data, which describes circuit behavior and data flow. In the same announcement, OpenAI also evaluated how far AI models can improve circuits using publicly available CPU cores.
AMD adopts AI-proposed changes in a GPU design sent to manufacturing
According to OpenAI's November 18 research announcement, in the collaboration with AMD, AI models modified RTL written in SystemVerilog and verified circuit behavior using Synopsys design software. The process checked that required functionality was preserved after each change and evaluated the effect on power, performance and area. AMD adopted some of the proposed changes in a design it recently sent to manufacturing, and incorporated other changes into design files for future products.
OpenAI has not disclosed the individual changes in power, performance and area behind the reported improvement of up to 14% on the GPU, how the improvement rate was calculated, or the average. It is also unclear whether the proposal that showed the largest improvement is the same one adopted in the design sent to manufacturing. The name of the product that adopted the changes and performance measured on actual chips have not been disclosed either.
OpenAI assessed the GPU optimization as achieving, in under five months, an improvement equivalent to roughly a year of progress under normal circumstances. AMD also reportedly confirmed single-digit PPA improvements in key circuit blocks of its next-generation flagship CPU. The results presented in this announcement sit at different verification stages: at AMD, AI-proposed changes were adopted in a design handed off to manufacturing; for Jalapeño, changes were confirmed when the chip incorporating them was brought up; and for RISC-V, they were evaluated with simulation and logic synthesis tools.
Jalapeño: a smaller multiplier and processing cut to one cycle
OpenAI's team estimates that AI model proposals reduced standard-cell area by about 13% in the vector unit and about 10% in the matrix unit of its inference chip, Jalapeño. Logic synthesis converts circuits described in RTL into standard cells, which are pre-built logic circuit components. Standard-cell area is the total area of the cells used, and the reductions are measured against each unit's area before the changes.
AI-assisted design began in early 2025 using GPT-4o. For the vector unit, the AI searched for improvements covering floating-point addition and multiplication, numerical conversion, operations that aggregate multiple values, and special functions. It also considered how many stages to divide operations into, looking for places where sharing circuitry and reducing the storage circuits that hold intermediate results could shrink the area.
For one of the multipliers in the matrix unit, cell area at logic synthesis was cut by about 15%, and the processing needed to produce a result was shortened from two cycles to one. For a separate adder, area was reduced by about 15% while keeping the two cycles needed to produce a result. Changes were also made to low-precision dot-product circuits and to the configuration of the pipeline, which splits processing into multiple stages.
However, fewer processing cycles does not automatically mean shorter execution time. The length of one cycle is determined by the clock frequency, and while reducing the number of stages can eliminate circuits that hold intermediate results, it can also increase the work per stage and lengthen signal propagation time. Moreover, as explained in Synopsys's overview of physical synthesis, real circuits must also account for delays from cell placement and wiring. How much a difference in cycle count affects execution time therefore depends on the final operating frequency.
OpenAI says it confirmed these optimizations when the actual chip was brought up. It has not, however, provided measured values showing what share of the total chip area each unit occupies or isolating the effect of individual changes. In its August 25 report on Jalapeño, the company said it completed the process from initial design to tape-out, the handoff of design data to manufacturing, in nine months. This announcement clarifies which arithmetic circuits the AI helped improve.
RISC-V evaluation: designing from a specification beats modifying existing RTL
To evaluate AI models' circuit design ability, OpenAI also ran experiments using the 32-bit embedded RISC-V CPU cores PicoRV32 and Ibex. There were two kinds of tasks: modifying existing RTL, and designing from scratch from a specification. For the latter, the model was given a code template for the external interfaces and written requirements, and left to decide the datapath and pipeline configuration. Tests to verify correct operation, memory and cache settings, and instruction throughput targets were set under common conditions for each core.
Scoring multiplied the geometric mean of IPC (instructions executed per clock cycle), measured on each workload not disclosed to the models, by the maximum clock frequency estimated by the Yosys logic synthesis tool. This estimates the number of instructions that can be executed per second. Candidates that failed simulation-based verification, or whose circuit area exceeded 16 times that of the original design, received zero. For each core, the highest instruction throughput observed under the same constraints was set to 1.0, and models were compared on average scores that included zero-scored attempts.
With a budget of about 112,000 tokens, GPT-6.2 Astra recorded the highest average score on both cores for optimizing existing RTL. On PicoRV32, its score was about double that of GPT-6 Astra. At comparable token budgets, GPT-6.1 Sol also outperformed GPT-5.6 Sol on both cores.
On the task of designing from a specification, by contrast, all five models evaluated scored higher on both cores than on modifying existing RTL. GPT-6.2 Astra's average score exceeded 0.7 on PicoRV32 and reached about 0.75 on Ibex. These scores, however, reflect not only the throughput of designs that passed verification but also the share of attempts that failed verification and scored zero.
The experiment as a whole comprised 9,881 trials, covering the two task types, the two CPU cores and all token budgets. The results come from an internal evaluation by OpenAI, and detailed conditions such as the full design data, the workloads and the technology library used have not been released. The evaluation is also based on simulation and logic synthesis tools, not measurements on actual chips.
Interpreting these results also requires considering the scoring criteria and the original design goals of the CPU cores. According to the README published by PicoRV32's developers, the core is designed to emphasize small circuit area and a high maximum clock frequency. Because this evaluation scores instructions executed per second, designs that add circuitry within the area limit and process instructions in fewer cycles could also earn high scores.
OpenAI itself acknowledges that in the design-from-specification task, being able to omit existing performance counters and additional instructions may have contributed to the score gap. For example, Ibex has a mechanism for configuring performance counters that record things such as memory read/write and multiplication wait times. According to Ibex's developer documentation, these help identify whether memory or computation is the cause when instruction execution slows. If the benchmark programs used for scoring do not use such features, omitting them may not affect the score. Which candidates actually omitted which features has not been disclosed.
Automating AI circuit design and the verification needed before manufacturing
On September 30, OpenAI and Synopsys announced they would jointly develop "GPT-Synopsys," an AI model for chip design. In the planned approach, AI models operate design and verification tools and repeatedly revise designs based on the analysis results, with human designers checking the final output. In the AMD collaboration too, the AI models verified their proposed changes using existing design tools. If generating and verifying candidates can be repeated at scale, designers gain more circuit configurations to compare and consider.
To evaluate the effect in real product development, it will be necessary to examine what share of AI-proposed changes pass functional verification, and which PPA improvements survive after placement and routing. If development effort can be reduced, including the time designers spend reviewing and correcting proposals, that time can be spent improving other circuits. In the AMD effort, the AI has progressed to finding changes that can be sent to manufacturing from a design that had already been optimized. Going forward, how much the adopted changes contribute to the power consumption and performance of actual chips will be the basis for judging the results.
