OpenAI's AI Improved an AMD Chip by Up to 14% in Power, Speed and Area in Five Months
Engineers spend months squeezing percentage points out of a finished chip design: less power, higher speed, less silicon. OpenAI's models improved these three metrics (PPA) in blocks of AMD's flagship GPU by up to 14% combined. According to OpenAI, the work took under five months instead of a year, and Synopsys software checked the changes. The takeaway: AI works for refining circuit designs when the result can be verified automatically.
AI-processed from 3DNews AI; edited by Hamidun News
Imagine you have working code that experienced people have already optimized, and you need to squeeze out a few more percent. Chips are the same, only more expensive: every percent saved in power, speed or silicon area means money across millions of chips. OpenAI said it had assigned this kind of refinement to its models. On blocks of AMD's flagship graphics processor, they improved three metrics at once: power consumption, performance and occupied area, which the industry abbreviates as PPA (Power, Performance, Area). The combined gain reached up to 14% relative to designs that engineers had already carefully optimized. According to OpenAI's estimate, it took less than five months, whereas the conventional approach would have required about a year.
What the AI Actually Did
The work began in early 2025 on OpenAI's own accelerator called Jalapeño, using the GPT-o3 mini model. The model worked on the vector block, the part of the chip that performs parallel computations. It optimized adders and multipliers, data conversion blocks and special-operation units. The model simplified logic circuits and redistributed computations across pipeline stages. As a result, different operations began sharing the same hardware components, and the amount of data temporarily stored in memory decreased.
The numbers are as follows. The multiplier takes up 15% less area and performs the operation in one cycle instead of two. The adder also shrank by 15%, while its latency stayed the same at two cycles. Overall, the standard-cell area of the vector block decreased by an estimated 13%, and that of the matrix block by 10%.
How It Worked with AMD
In cooperation with AMD, OpenAI's models proposed changes to RTL code in SystemVerilog, the language used to describe digital circuits at the hardware level. The model looked not only at the computational blocks but also at the control and data-transfer circuits.
The AI's suggestions were not taken on faith. The circuits were checked with Synopsys tools, which confirmed that the blocks retained their original behavior after the changes and helped evaluate trade-offs between power consumption, performance and die area. AMD has already included some of the changes in one design sent to production, and added the rest to the design documentation for future chip versions. In addition, the company confirmed a PPA gain of several percent in key blocks of next-generation flagship central processors designed with OpenAI's models.
Test on Open RISC-V Cores and Caveats
To compare model generations, OpenAI took two open RISC-V processor cores: PicoRV32, designed for programmable chips, and Ibex, which is used in OpenTitan. There were two tasks: speed up an existing core and redesign a processor from a requirements description. With a comparable amount of computation — about 112 thousand tokens per run — GPT-6.2 Astra showed the best result on both cores, and GPT-6.1 Sol outperformed GPT-5.6 Sol. In total, 9,881 trials were conducted, and all five models received higher scores when creating a processor from scratch than when improving an existing circuit.
This advantage has a downside, which OpenAI itself warns about. When designing from scratch, one can leave out features of the original processor that barely affect the results of the specified tests. Therefore, such tests do not fully reflect real-world development. For a rigorous evaluation, mandatory features, tighter chip-area constraints and circuit verification after physical design are needed.
What This Means in Practice
The main lesson for those building agents: AI delivers results where automatic verification exists. Here, Synopsys tools provided it, and the pairing of "the model proposes, verification software confirms" turns ideas into changes fit for production. The same applies to ordinary development: an agent equipped with tests and linters can be trusted with refining finished code far more readily than one without them.
OpenAI's next step is with Synopsys: the companies are developing a GPT-Synopsys model that will be able to use electronic design automation and circuit verification tools the way experienced chip developers do.
The limitations are worth keeping in mind. All the figures are OpenAI's own estimates; there are no independent measurements. The 14% figure is an "up to" number, not an average across all blocks. The comparison of "a year versus five months" is an estimate, not a controlled experiment. But the fact that some of the changes made it into a design sent to production shows that AI in chip design has already moved beyond demonstrations.
- Nov 21, 2026Anthropic goes public and raises a record $107.3 billion
- Nov 21, 2026OpenAI launches ChatGPT Wallet: agents buy without ever seeing your card number
- Nov 21, 2026Grok Build left 35,578 apps open — 94.7% of live projects require no login
- Nov 21, 2026OpenAI to Expand Academic Program to 200,000 Seats, but Swaps Pro for Plus
Key Questions Answered
What the AI Actually Did?
The work began in early 2025 on OpenAI's own accelerator called Jalapeño, using the GPT-o3 mini model. The model worked on the vector block, the part of the chip that performs parallel computations. It optimized adders and multipliers, data conversion blocks and special-operation units. The model simplified logic circuits and redistributed computations across pipeline stages. As a result, different operations began sharing the same hardware components, and the amount of data temporarily stored in memory decreased.
How It Worked with AMD?
In cooperation with AMD, OpenAI's models proposed changes to RTL code in SystemVerilog, the language used to describe digital circuits at the hardware level. The model looked not only at the computational blocks but also at the control and data-transfer circuits. The AI's suggestions were not taken on faith.
What are the key details on Test on Open RISC-V Cores and Caveats?
To compare model generations, OpenAI took two open RISC-V processor cores: PicoRV32, designed for programmable chips, and Ibex, which is used in OpenTitan. There were two tasks: speed up an existing core and redesign a processor from a requirements description. With a comparable amount of computation — about 112 thousand tokens per run — GPT-6.2 Astra showed the best result on both cores, and GPT-6.1 Sol outperformed GPT-5.6 Sol.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.