On November 18, OpenAI published a chip design report saying its models contributed to two kinds of hardware work: Jalapeño, its in-house inference chip, and the design of modules in a flagship AMD GPU, developed together with AMD. According to the company, on AMD's flagship GPU IP the changes proposed by the models improved a combined power, performance and area metric (collectively known in the industry as PPA) by up to 14% against a "highly optimized baseline." The work took less than five months, and some of the changes are already in designs sent to the fab for manufacturing.
OpenAI sums up the result as "about a year of progress in under five months." That is the company's own framing; in public reporting, AMD has not given a separate quantified statement.
What the models changed is RTL code
According to media accounts of the report, the models rewrote RTL, the code that describes a chip's circuits, written in SystemVerilog. Every rewrite was checked with Synopsys tools to confirm that module behavior had not changed, and only then were the PPA trade-offs evaluated. Beyond the GPU design headed to the fab, AMD also placed another set of changes into the production design files of a later version, and validated single-digit percentage PPA improvements on several key modules of a next-generation flagship CPU.
The Jalapeño figures are more detailed. Changes proposed by the models cut the standard-cell area of the warp engine (the vector unit) by about 13% and that of the matrix unit by about 10%. Within the matrix unit, a redesigned multiplier used roughly 15% less cell area at synthesis and dropped its latency from two cycles to one; the adder also used about 15% less area while keeping a two-cycle latency. OpenAI says these optimizations were confirmed at chip bring-up.
The report also includes a set of RISC-V experiments. Five generations of models were asked to optimize two open-source 32-bit processor cores, PicoRV32 and Ibex, or to write an entire core directly from written requirements. There were 9,881 scored runs in total, and any run that failed simulation or came out more than 16 times larger than the original core scored zero. GPT-6.2 Astra had the highest average score on both cores in the RTL optimization task; in the "write a core from requirements" task, its average score exceeded 0.7 on PicoRV32 and was about 0.75 on Ibex.
OpenAI also spelled out the flaws in these experiments. Models working from requirements sometimes left out features present in the original cores, such as performance counters and extension instructions, which do little for throughput on the test workloads. That means the "core from requirements" scores may be inflated in part by missing functionality. The company's proposed fix is to make those features mandatory, tighten the area cap, and re-check timing after physical synthesis.
Connecting the dots
Looking at OpenAI's hardware moves over the past year, the pace has been rapid. In June it announced Jalapeño with Broadcom, saying the chip took about nine months from initial design to tape-out. In August it released silicon results. On September 30 it signed a multi-year agreement with Synopsys covering joint development, go-to-market and revenue sharing. In this report, the two companies went further and announced that they will train GPT-Synopsys, a model meant to operate EDA tools the way a senior engineer would, carrying work from PPA optimization through timing and verification closure, with engineers reviewing at the final stage.
Synopsys says customer design data will not be used to train GPT-Synopsys, that data is encrypted at rest and in transit, and that retention periods, auditing and permissions are configurable. Design data is among the most sensitive assets a chip company has, and whether this pledge eases concerns will depend on how many customers are willing to hand their RTL to the tool.
For now, every number comes from OpenAI alone. Which modules the 14% applies to, and which AMD product generation it corresponds to, has not been disclosed in public materials, and no third party has reproduced the result.
Sources: OpenAI chip design report, The Decoder, CocoLoop, Unite.AI; figures checked for the PPA gain on AMD flagship GPU IP, Jalapeño per-unit area reductions and the number of RISC-V experiment runs.