OpenAI Unveils AI-Assisted Chip Design Results: Up to 14% PPA Improvement in Key AMD IP Blocks
OpenAI has released a technical report revealing that its models reduced the standard-cell area of the vector and matrix units in its custom Jalapeño AI inference chip by approximately 13% and 10%, respectively, while improving PPA metrics for some key IP blocks in AMD's flagship GPU by up to 14%. Across 9,881 scored runs, GPT-6.2 Astra achieved the highest average score on RTL optimization tasks for PicoRV32 and Ibex. OpenAI also announced a partnership with Synopsys to train GPT-Synopsys for chip design and verification workflows.
Why read this
SelectedAI score 77
The report quantifies the area and PPA gains achieved by AI-optimized chip RTL, while also highlighting capability differences between optimizing existing designs and generating designs from specifications.
Topics
Body · AI translation
AI Begins Optimizing Its Own “Brain”: OpenAI Releases Latest Chip Design Results
Phoenix Technology — Download the app for an exclusive first look November 19, 2026, 06:40:04, from Beijing Updated November 19, 2026, 06:41:24
Phoenix Technology News — On November 19 Beijing time, OpenAI released a new technical report, “Advancing chip design with OpenAI models,” detailing its latest progress in using artificial intelligence models to assist chip design. OpenAI said its models not only helped optimize circuit designs for Jalapeño, its internally developed AI inference chip, but also delivered improvements of up to 14% in power, performance, and area (PPA) metrics for key IP blocks in AMD’s flagship GPUs through a partnership with AMD. AMD has incorporated some of these optimizations into chip designs recently submitted for fabrication.
OpenAI said AI models are moving beyond assisting engineers with code and entering the underlying circuit design and microarchitecture optimization stages of chip development. The results announced this time cover OpenAI’s own AI chips, AMD’s flagship GPUs and next-generation CPUs, as well as open-source RISC-V processor designs.
For its internally developed chip, OpenAI began using GPT-4o in the design of Jalapeño in early 2025. Even that relatively early model was already capable of identifying opportunities for optimization in conventional chip design workflows.
Specifically, in Jalapeño’s vector compute unit, the AI model reduced hardware resource usage and intermediate-data storage requirements by adjusting floating-point adders, multipliers, data-conversion logic, and the pipeline architecture. In the matrix compute unit, the model redesigned circuits for floating-point operations and low-precision dot products.
One multiplier design optimization reduced the synthesized standard-cell area by about 15% while shortening operation latency from two clock cycles to one. Another adder optimization likewise achieved an area reduction of about 15% while retaining a latency of two clock cycles.
OpenAI estimates that the design changes proposed by its AI models ultimately reduced the standard-cell area of Jalapeño’s vector unit by about 13% and that of its matrix unit by about 10%. These improvements have been confirmed during silicon bring-up and validation.
Notably, OpenAI is now applying these capabilities to AMD’s commercial chip designs.
According to information disclosed by OpenAI, the two companies worked together to optimize key IP blocks in AMD’s flagship GPUs. Starting from designs that had already been highly optimized, OpenAI’s models further improved power, performance, and area metrics by rewriting the SystemVerilog RTL code that describes the chip circuitry, with PPA improvements reaching as high as 14% for some designs.
OpenAI said this was equivalent to achieving in less than five months design progress that would ordinarily take about a year. AMD has applied some of the optimization proposals to chip designs recently submitted for fabrication, while other improvements have been incorporated into production design files for use in future versions.
AMD has also validated optimizations produced by OpenAI’s models for key blocks in its next-generation flagship CPUs, with the relevant designs achieving single-digit percentage improvements in PPA metrics.
OpenAI said its models can now analyze RTL code in a manner approaching that of experienced microarchitecture engineers and propose modifications to compute logic, control circuitry, and data-transfer structures. The models also use Synopsys electronic design automation tools to verify that modifications preserve the original functionality and to assess trade-offs among power, performance, and area in different designs.
In addition to real-world chip projects, OpenAI evaluated the chip-design capabilities of different generations of models using two open-source 32-bit RISC-V processor cores, PicoRV32 and Ibex.
The tests included two categories of tasks: optimizing processor throughput based on an existing RTL design, and generating a processor RTL design from scratch based on written specifications.
The results showed that, under a comparable budget of about 112,000 Tokens per run, GPT-6.2 Astra achieved the highest average score on RTL optimization tasks for both processor cores. In the PicoRV32 test, its average score was about twice that of GPT-6 Astra.
More notably, in tasks involving the direct generation of RTL designs from written specifications, all five models tested scored higher on both processors than they did when optimizing existing designs. GPT-6.2 Astra achieved an average normalized score above 0.7 in the PicoRV32 test and about 0.75 in the Ibex test.
The evaluation comprised a total of 9,881 scored runs spanning different models, tasks, processor cores, and Token budgets.
OpenAI cautioned, however, that these tests still have limitations. When redesigning a processor from specifications, a model can omit features such as performance counters or certain additional instructions, so a higher score does not necessarily mean the new design would offer equivalent advantages in a real-world product. Future evaluations will still need to incorporate stricter functional requirements, chip-area constraints, and post-physical-synthesis timing checks.
In the next phase, OpenAI announced a partnership with chip-design software vendor Synopsys to train a new model called GPT-Synopsys, enabling it to use electronic design automation (EDA) tools like a professional chip-design engineer and further expanding AI’s role in chip design and verification workflows.
The results disclosed this time indicate that OpenAI is attempting to establish a technological pathway in which AI is used to optimize chips, which in turn support more efficient operation of AI models. For now, however, the main results announced remain focused on specific compute blocks and evaluation tasks and cannot be directly equated with an entire GPU or CPU achieving performance gains of the same magnitude. (Editor: Yu Lei)
For more firsthand news, download the Phoenix News app and subscribe to Phoenix Technology. For in-depth reporting, search WeChat for “Phoenix Technology.”
AI Trends