OpenAI Uses AI to Optimize Chip Design, Completing About a Year’s Work in Five Months and Improving PPA by Up to 14%
OpenAI announced that in under five months, its models improved the PPA of AMD’s flagship GPU IP by up to 14%, equivalent to about a full year of R&D work by human engineers. Some changes have already been incorporated into recent tape-outs and subsequent mass-production designs. The standard-cell areas of Jalapeño’s vector and matrix units were reduced by approximately 13% and 10%, respectively. In benchmarks involving two open-source 32-bit RISC-V cores, GPT‑6.2 Astra achieved the highest average score, and all five models scored higher when designing processor cores from written requirements than when optimizing existing designs.
Why read this
SelectedAI score 83
The materials also present optimization results for mass-produced chips and benchmark results for open-source RISC-V designs, enabling comparisons of the models’ capabilities in incremental optimization and greenfield design.
Topics
Body · AI translation
Zhidx reported on November 19 that OpenAI has just announced its models improved the PPA—a core set of chip-design metrics—of AMD’s flagship GPU IP by as much as 14% in less than five months, equivalent to roughly a full year of R&D work by human engineers.
OpenAI’s research demonstrated that, in RISC-V evaluation tasks, its models could explore entirely new processor architectures and deliver incremental optimizations to mature, production-ready designs such as Jalapeño and AMD’s next-generation flagship GPUs and CPUs.
For Jalapeño, OpenAI’s first in-house inference chip, model-proposed modifications reduced the standard-cell area of the vector unit by approximately 13% and that of the matrix unit by around 10%. OpenAI said these improvements had been validated during Jalapeño’s silicon bring-up and debugging phase.
To measure its models’ capabilities in end-to-end chip-design tasks, OpenAI evaluated successive generations of models on two open-source processor designs. The results showed that all five models—GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol, and GPT-5.5—scored higher when designing processor cores from scratch based on written requirements than when optimizing existing designs.
Meanwhile, to ensure that capability improvements are grounded in a trustworthy evaluation system, OpenAI is collaborating with Synopsys to integrate frontier AI capabilities into the EDA toolchains engineers currently use for chip design and verification.
In early 2025, GPT‑4o was already being used to help design Jalapeño, OpenAI’s in-house inference chip. That early model achieved several substantial PPA optimizations.
The vector unit is responsible for scheduling and executing parallel computing tasks. Improvements to this module came from the floating-point adders, multipliers, conversion circuits, reduction logic, and special-function units. The model explored multiple logic implementations while studying how to divide computational tasks across different pipeline stages, identifying opportunities to reuse hardware and reduce intermediate-storage overhead.
The matrix unit is a computational component whose gains came from improvements to floating-point adders, multipliers, low-precision dot-product circuits, and other elements. One multiplier was redesigned, reducing its synthesized cell area by approximately 15% while cutting latency from two clock cycles to one. A revised adder likewise achieved an approximately 15% reduction in cell area while retaining its two-cycle latency. Further optimization of the dot-product logic and pipeline structure produced additional gains for the module.
OpenAI collaborated with AMD on PPA optimization, achieving gains of up to 14% for flagship GPU IP over an already heavily optimized baseline.
OpenAI said the model uncovered additional optimization opportunities by rewriting the code that describes chip circuitry. This was equivalent to completing roughly a full year of R&D work in less than five months. AMD has incorporated some of the optimized changes into designs recently sent to the foundry for tape-out, while other optimizations have been added to production design files for subsequent iterations.
Researchers observed that the model could reason about and analyze RTL code at the level of a senior microarchitect. It could rewrite circuits responsible for computation, control operations, and data movement. Using Synopsys tools, it verified that each rewrite preserved the module’s original functionality and assessed the trade-offs among power, performance, and area.
AMD’s verification showed that designs generated using OpenAI’s models delivered PPA improvements of several percentage points across multiple key modules of its next-generation flagship CPU.
The results OpenAI achieved on Jalapeño and AMD chips demonstrate that its models can optimize production-grade designs.
To measure progress in chip-design capabilities across model generations, OpenAI tested them on two open-source 32-bit RISC-V cores: PicoRV32, a streamlined core intended for reprogrammable chips such as FPGAs, and Ibex, which OpenTitan uses as the embedded processor in its security chip. Both are complete processor cores designed for embedded and microcontroller-class workloads.
Researchers provided each model generation with the original hardware-description code for each processor core and asked it to improve throughput. In other words, a modification might reduce the number of clock cycles required for an instruction while increasing the duration of each individual clock cycle; the throughput metric captures the combined effect of both changes.
They verified the output of every design through simulation. A design received a score of 0 if it failed verification or occupied more than 16 times the area of the original core. For designs that passed verification, they used test workloads the models had never seen to measure instructions per clock cycle (IPC). They then took the geometric mean of the workload results and multiplied it by the maximum clock frequency estimated by Yosys to obtain an estimated number of instructions per second.
For each processor core, they normalized the best throughput observed under the same constraints to a baseline score of 1.0.
With an equal resource budget of approximately 112,000 tokens per call, GPT‑6.2 Astra achieved the highest average score on both processor cores, while GPT‑6.1 Sol outscored GPT‑5.6 Sol on both cores.
In the second task, researchers provided the models with written requirements and templates from the open-source code. These templates defined how the processor core connected to the rest of the system, while the model independently chose how data would flow within the core and how work would be divided across multiple pipeline stages.
For both processor cores, correctness verification, memory and cache configurations, and throughput targets were kept consistent with the preceding RTL optimization task.
With a budget of approximately 112,000 tokens per call, all five models scored higher when designing processor cores from scratch based on written requirements than when optimizing existing designs. This finding held for both the PicoRV32 and Ibex cores.
GPT‑6.2 Astra’s improvement was greater on Ibex than on PicoRV32. The full evaluation comprised 9,881 valid scored runs covering both tasks, both processor cores, and all token-budget configurations.
When designing from scratch based on written requirements, the models could choose to omit certain features included in the original cores, such as performance counters and extended instructions. Such features contributed little to throughput on the test workloads used here, which may partly explain the higher scores on this task.
OpenAI said the evaluation would more closely resemble real-world production chip-design scenarios if it required these features to be retained, imposed stricter circuit-area constraints, and included timing checks after physical synthesis.
03.
OpenAI’s research results show that, in chip design, AI is already capable of uncovering incremental gains in mature, production-ready designs and independently exploring the microarchitecture space based on specifications.
This could create a new battleground for future chip competition. Chipmakers will need to focus not only on fabrication processes, packaging, and architecture itself, but also on using AI to improve efficiency, converge on PPA targets more quickly, and shorten R&D cycles.
Source: OpenAI
This article is from the WeChat public account “Zhidx.” Author: Cheng Qian; editor: Li Shuiqing. Published by 36Kr with authorization.
AI Trends