Abstract:
OpenAI released its latest technical report "Advancing chip design with OpenAI models", disclosing its latest progress in using artificial intelligence models to assist chip design. OpenAI stated that its model not only helped the self-developed AI inference chip Jalapeño optimize the circuit design, but also achieved up to 14% improvement in power consumption, performance and area (PPA) indicators of key IP modules of the flagship GPU in cooperation with AMD
. Some of the optimization results have been incorporated into chip designs recently submitted for manufacturing by AMD.

OpenAI stated that the AI model is moving from assisting engineers in writing code to further entering the chip's underlying circuit design and microarchitecture optimization. The results announced this time cover self-developed AI chips, AMD flagship GPUs and next-generation CPUs, as well as open source RISC-V processor designs.
In terms of self-developed chips, OpenAI will use GPT-4o to participate in the design of Jalapeño starting in early 2025. Even the earlier models at the time could already find room for optimization in the traditional chip design process.
Specifically, in Jalapeño’s vector computing unit, the AI model reduces hardware resource usage and intermediate data storage requirements by adjusting floating-point adders, multipliers, data conversion logic and pipeline structures. In the matrix computing unit, the model is redesigned for floating point operations and low-precision dot product circuits.
Among them, a multiplier design optimization reduces the area of the standard unit after synthesis by approximately
15%
, while reducing the operation delay from two clock cycles to one clock cycle. Another adder optimization also achieves an area reduction of about 15% while maintaining a two-clock cycle delay.OpenAI estimates that the design modifications proposed by the AI model ultimately reduced the standard unit area of the Jalapeño vector unit by approximately
13%
, the matrix unit is reduced by about10%
. These improvements have been confirmed during silicon debug verification of the chip.It is worth noting that OpenAI is applying the above capabilities to AMD's commercial chip design.
According to information disclosed by OpenAI, the two parties have cooperated to optimize key IP modules of AMD’s flagship GPU. Based on the original highly optimized design, the OpenAI model further improves power consumption, performance and area indicators by rewriting the SystemVerilog RTL code that describes the chip circuit. The PPA improvement of some designs reaches up to
14%
.OpenAI said this is equivalent to completing design progress that would normally take about a year in less than five months.
AMD has applied some optimizations to chip designs recently submitted for manufacturing, and other improvements have been incorporated into production design documents and are planned for future versions.At the same time, AMD also verified the optimization results of the OpenAI model in key modules of the next-generation flagship CPU, and the related design achieved a single-digit percentage improvement in PPA indicators.
OpenAI stated that its model has been able to analyze RTL code in a manner close to that of a senior microarchitecture engineer, and propose modifications to computing logic, control circuits and data transmission structures. The model also leverages Synopsys' electronic design automation tools to verify that modifications maintain original functionality and evaluate the power, performance and area trade-offs of different designs.
In addition to the actual chip project, OpenAI also evaluated the chip design capabilities of different generation models through two open source 32-bit RISC-V processor cores PicoRV32 and Ibex.
The test includes two types of tasks: one is to optimize the processor throughput based on the existing RTL design, and the other is to generate the processor RTL design from scratch according to the text specification requirements.
The results show that with a comparable budget of approximately 112,000 Tokens per run,
GPT-6.2 Astra achieved the highest average score in the RTL optimization tasks of both processor cores
. Among them, in the PicoRV32 test, its average score is about twice that of GPT-6 Astra.More notably, in the task of generating RTL designs directly from text specifications, all five tested models scored higher on both processors than in the task of optimizing existing designs. GPT-6.2 Astra achieved an average normalized score of over 0.7 in the PicoRV32 test and around 0.75 in the Ibex test.
This evaluation included a total of
9881 scoring runs
, covering different models, tasks, processor cores and token budgets.However, OpenAI also pointed out that these tests still have limitations. Since models can omit features such as performance counters or some extra instructions when redesigning a processor from specs, a higher score does not necessarily mean that the new design will be equally advantageous in actual products. In the future, more stringent functional requirements, chip area restrictions, and timing checks after physical synthesis will still need to be added.
In the next stage, OpenAI announced its cooperation with chip design software manufacturer Synopsys, and the training name is
GPT-Synopsys
The new model enables it to use electronic design automation (EDA) tools like professional chip design engineers, further improving the application capabilities of AI in the chip design and verification process.Judging from the results disclosed this time, OpenAI is trying to establish a set of technical paths that use AI to optimize chips and then use more efficient chips to support the operation of AI models. However, the main results announced so far are still focused on specific computing modules and evaluation tasks, and cannot be directly equated to the same proportion of performance improvement for the entire GPU or CPU.
Comments