OpenAI Reports Up to 14% PPA Gains in AMD Chip Designs
OpenAI reports that model-written chip RTL delivered up to 14% PPA gains in AMD GPU work, reduced Jalapeño area and improved blocks in an upcoming flagship CPU.

Why it matters
- OpenAI reported up to 14% PPA gains on AMD flagship GPU IP on November 18, 2026.
- GPT-4o changes cut standard-cell area by about 13% in Jalapeño’s warp engine and 10% in its matrix unit.
- AI assistance helped Jalapeño move from initial design to tapeout in nine months.
- GPT-6.2 Astra led both RISC-V cores in a 9,881-run evaluation at roughly 112,000 tokens per run.
- The GPT-Synopsys partnership builds on a multi-year agreement announced on September 30, 2026.
By Sophie Lindqvist
OpenAI reported on November 18, 2026 that its models achieved up to a 14% improvement in power, performance and area on heavily optimized AMD flagship GPU intellectual property. Some model-proposed circuit rewrites were included in a GPU design sent for manufacturing, while AMD added other changes to production design files for future versions.
The models rewrote SystemVerilog RTL, the code that describes a chip’s circuits. OpenAI said the models used Synopsys tools to verify that each rewrite preserved a block’s behavior and assess PPA trade-offs. That process links the claimed gains to established chip-design checks instead of treating generated code as a finished design.
OpenAI described the GPU effort as “roughly a year of progress achieved in less than five months.” AMD verified single-digit percentage PPA gains across key blocks of an upcoming next-generation flagship CPU that used model-written designs.
The two programs reached different validation stages. OpenAI’s GPU work included a design sent for manufacturing, while AMD’s CPU result covered verified gains in selected blocks of a future processor.
How much did GPT-4o shrink OpenAI’s Jalapeño chip?
OpenAI began using GPT-4o to help design Jalapeño, its first custom inference chip, in early 2025. The team estimates that model-proposed changes reduced standard-cell area by about 13% in the warp engine, which manages and executes parallel computing tasks. In the matrix unit, an arithmetic component, the estimated reduction was 10%.
OpenAI said the models explored four kinds of improvement:
- Alternative logic implementations.
- Ways to divide computations across pipeline stages.
- Opportunities to share hardware.
- Reductions in intermediate storage.
Each option targeted a different cost in the block: logic structure, pipeline balance, hardware duplication or temporary data storage. One matrix-unit multiplier redesign used about 15% less cell area in synthesis while reducing latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% while retaining two-cycle latency.
OpenAI said changes to dot-product logic and pipeline organization contributed additional savings across the block. Its August 25, 2026 silicon results said AI assistance helped Jalapeño move “from initial design to tapeout in nine months.” The company said the optimizations were proven in silicon during the chip’s bring-up.
The matrix multiplier changed both area and latency, while the adder saved area without adding a cycle. Together, the examples show the models changing circuit structure under explicit verification rather than producing code without checks.
In plain terms, GPT-4o contributed measurable area and latency reductions, with the reported optimizations validated in silicon.
Which models led OpenAI’s RISC-V evaluation?
OpenAI tested GPT-6.2 Astra, GPT-6 Astra, GPT-6.1 Sol, GPT-5.6 Sol and GPT-5.5 on two open-source 32-bit RISC-V cores. PicoRV32 is a compact core designed for reprogrammable FPGAs, while OpenTitan uses Ibex as the embedded processor in its security chip.
OpenAI counted 9,881 scored runs across the tasks, cores and token budgets. The evaluation covered two jobs:
- Optimizing an existing core’s RTL.
- Creating a core from written requirements and interface templates.
The two tasks test different problems. RTL optimization starts with a working core and asks for a faster implementation, while specification-to-RTL asks the model to produce the core itself.
For RTL optimization, each model received the original hardware description and was asked to increase throughput, the number of instructions executed per second. A design scored zero if it failed a simulation check or used more than 16 times the original core’s area.
For passing designs, OpenAI took the geometric mean of instructions per clock across workloads hidden from the models. It multiplied that value by the maximum clock frequency estimated with Yosys.
The best observed throughput for each core under the same constraints received a score of 1.0, and averages included zero-scoring runs. Because each core used its own reference point, 1.0 marked the benchmark ceiling rather than a universal performance rate. A zero-scoring run therefore reduced the final average.
At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores in RTL optimization. GPT-6.1 Sol also outscored GPT-5.6 Sol on both.
In specification-to-RTL design near the same budget, GPT-6.2 Astra recorded a mean normalized score above 0.7 on PicoRV32 and averaged about 0.75 on Ibex. All five models scored higher when designing from written requirements than when optimizing existing RTL on both cores.
GPT-6.2 Astra led both cores, but written requirements produced higher scores for every model.
Why do the specification scores need caution?
OpenAI said models starting from written requirements could omit features found in the original cores, including performance counters and additional instructions. Those omitted features did little to improve throughput on the test workloads, and their absence could explain part of the score advantage.
OpenAI said a more representative production test would need to require those features, set a tighter circuit-area limit and check timing after physical synthesis. The company’s qualification matters because higher scores can reward a design that leaves low-throughput features out.
The distinction makes specification-based scores a directional signal about code generation, not a direct measure of full-feature optimization. OpenAI’s caveat is material because the task can reward leaving out requirements that the test workloads barely exercise.
The higher specification scores are useful signals, but OpenAI says they are not yet production-equivalent measurements.
What will GPT-Synopsys do inside chip-design workflows?
OpenAI is partnering with Synopsys to train GPT-Synopsys to use electronic design automation tools the way expert chip designers do. The work builds on a multi-year agreement announced on September 30, 2026 that covers joint development, go-to-market collaboration and revenue sharing.
Synopsys said engineers would delegate objectives from PPA optimization through timing and verification closure. In that setup, agents run tools, interpret results, implement changes and iterate toward verified outcomes for engineer review.
Synopsys said customer design data would not be used to train GPT-Synopsys. The data would be encrypted at rest and in transit, with configurable retention, audit and permission controls.
Synopsys said those protections, plus required engineer review, address the realities of working on proprietary production designs. OpenAI framed the central challenge as joining model-driven design gains to trusted verification, with GPT-Synopsys intended to connect circuit optimization and broader EDA workflows.
Required review is the final control in that workflow. It keeps agents inside a delegation model in which software proposes and executes changes, while engineers assess the verified outcome before it advances.
The partnership aims to turn code generation into a controlled design process backed by verification and human approval.
The next test is whether the reported gains survive full production workflows that preserve required features, pass physical synthesis and remain under engineer review. OpenAI’s benchmark limits and Synopsys’s proposed controls make that the practical measure of progress.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
777 articles
Related articles
- OpenAI says its models cut AMD flagship GPU IP power-performance gaps by 14%
- OpenAI models deliver up to 14% AMD PPA gain and Jalapeño area cuts
- OpenAI Commits to 6 Gigawatts of AMD GPUs in Multi-Year Deal
- OpenAI's Jalapeño Chip: LLMs Cut Design Time to Record Lows
- OpenAI Says Its Models Can Now Design Hardware End-to-End