OpenAI Reports Chip Design Gains at AMD and Jalapeño | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
OpenAI Reports Chip Design Gains at AMD and Jalapeño
OpenAI says model-written circuit changes reached AMD manufacturing files and working Jalapeño silicon, while its RISC-V benchmarks show progress under looser constraints.
OpenAI reports that its AI models successfully optimized circuit designs for AMD hardware and the company's own Jalapeño inference chip. By rewriting SystemVerilog register-transfer-level code, the models achieved up to 14 percent improvements in power, performance, and area for AMD flagship GPU intellectual property, with some changes entering manufacturing. For Jalapeño, the models reduced standard-cell area by 10 to 13 percent in critical warp and matrix units. These results are significant because they moved beyond simulation, proving successful during silicon bring-up, though they represent localized block optimizations rather than whole-chip performance gains.
OpenAI says its models have improved AMD chip designs enough for some changes to enter a design sent for manufacturing. It also says optimizations proposed by GPT-4o worked during silicon bring-up of Jalapeño, its first custom inference chip.
The company’s November 18, 2026 announcement reports up to 14% improvement in power, performance and area, or PPA, on AMD flagship GPU intellectual property. Separately, OpenAI estimates model-proposed changes reduced standard-cell area by about 13% in Jalapeño’s warp engine and 10% in its matrix unit.
Those results carry more engineering weight than a successful simulation alone. But the validation stages differ: AMD accepting a circuit change into a manufacturing design, Jalapeño completing bring-up, and an open-source processor earning a benchmark score answer different questions. The useful contribution here is to separate those milestones before comparing the percentages.
What Reached Production Designs or Silicon?
OpenAI describes five main design targets, spanning proprietary hardware and two open-source processor cores. This table separates the reported result from the evidence supporting it.
Design target
Reported result
Validation described
AMD flagship GPU IP
Up to 14% PPA improvement against heavily optimized baselines
Some changes entered a design sent for manufacturing; others entered production design files for future versions
AMD next-generation flagship CPU
Single-digit percentage PPA gains across key blocks
AMD verified the gains, according to OpenAI; no manufacturing milestone is specified
Jalapeño
About 13% less standard-cell area in the warp engine and 10% less in the matrix unit
OpenAI says the optimizations were proven during silicon bring-up
PicoRV32
GPT-6.2 Astra led the existing-core optimization comparison at comparable larger budgets
Simulation checks and estimated throughput, with scores normalized within the evaluation
Ibex
GPT-6.2 Astra had the highest average existing-core optimization score at about 112,000 tokens per run
Simulation checks and estimated throughput under the same general evaluation method
The figures and production milestones are company-reported. The supplied reporting contains no independent benchmark of the claimed gains, and the account of AMD’s decisions comes through OpenAI.
My view is that the strongest evidence is the acceptance of specific changes into a manufacturing design and their successful operation in Jalapeño silicon. Those are consequential milestones. They still leave questions about the size of any whole-chip benefit.
AMD’s 14% Claim Applies to Design IP
OpenAI says its model rewrote SystemVerilog register-transfer-level code, usually shortened to RTL, for AMD. RTL describes the circuit’s operations, registers and movement of data. Changing it can alter the hardware that design tools eventually produce.
The model used Synopsys tools to check that each rewrite preserved the block’s behavior and to evaluate its power, performance and area trade-offs. That distinction matters: plausible-looking circuit code is only a proposal until verification and implementation tools establish whether it works and offers a useful improvement.
PPA also combines three different engineering concerns:
Power: how much electrical power a circuit consumes under specified conditions.
Performance: how quickly it can complete work or meet timing requirements.
Area: how much implementation space its circuitry requires.
OpenAI does not provide enough detail in the announcement to treat “up to 14%” as a 14% improvement in each measure, or as a 14% speedup for a complete GPU. The reported result concerns flagship GPU IP relative to heavily optimized baselines. It supplies no application benchmark showing what a buyer would experience.
The manufacturing detail is more concrete. According to OpenAI’s account, AMD included some model-written changes in a recent design sent for manufacturing and placed others in production design files for future versions. A manufacturing submission is a meaningful commitment, but it is not evidence that the resulting chip has completed bring-up or reached customers.
OpenAI describes the work as “roughly a year of progress achieved in less than 5 months.” That is the company’s assessment of development progress, not a published measurement showing that an entire chip program ran more than twice as fast.
The CPU claim is narrower still: OpenAI says AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU. No corresponding manufacturing or silicon milestone is stated for those changes.
I’d read these results as evidence that models can find worthwhile local improvements even in mature designs. Extending them to a whole-chip performance forecast would require knowing which blocks changed, how much of the chip they occupy and what limits the workloads being measured.
Jalapeño Provides the Silicon Bring-Up Evidence
OpenAI began using GPT-4o to help design Jalapeño in early 2025. Its technical account gives concrete examples of what the model changed.
In the warp engine, which manages and executes parallel computing tasks, proposed optimizations covered floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit. The model explored alternative logic implementations and different ways to distribute calculations across pipeline stages.
Pipeline organization matters because a calculation can be split across several clock cycles. Changing those boundaries can affect intermediate storage, hardware sharing, latency and the timing requirements of each stage.
The matrix unit supplied two particularly useful examples:
A multiplier redesign used about 15% less cell area in synthesis and reduced latency from two clock cycles to one.
A separate adder redesign used about 15% less cell area while retaining its two-cycle latency.
These are distinct improvements. The multiplier result combines smaller synthesized circuitry with fewer cycles for that operation. The adder result saves area without changing its stated cycle latency. Neither number, by itself, gives a whole-chip speedup.
Across the broader design effort, OpenAI estimates that model-proposed changes reduced standard-cell area by approximately 13% in the warp engine and 10% in the matrix unit. Standard-cell area concerns the logic cells used to implement those blocks; it should not be read as an equivalent reduction in total chip die area.
OpenAI says the optimizations were proven during Jalapeño’s silicon bring-up, the stage when manufactured chips are powered up and checked. That adds evidence that the changes survived implementation and operated in hardware. The area percentages remain the team’s reported estimates, rather than independently published measurements of the finished chip.
The practical significance is the combination: specific circuit changes, quantified implementation benefits and successful bring-up. It is a stronger case than generated RTL that merely passes a testbench.
The RISC-V Tests Measure a Different Capability
OpenAI also evaluated successive models on two open-source 32-bit RISC-V processors. PicoRV32 is a compact core designed for FPGAs, while Ibex serves as the embedded processor in OpenTitan’s security chip.
These are complete processor cores for embedded, microcontroller-class workloads. They provide repeatable research targets, but their benchmark results should stay separate from the AMD and Jalapeño production claims.
OpenAI used two tasks:
Optimize an existing core: The model received the original hardware description and tried to increase throughput.
Design from requirements: The model received written requirements and interface templates, then chose the core’s internal organization.
For the second task, OpenAI kept the correctness checks, memory and cache settings, and throughput objectives the same as in the optimization task. That helps make the two exercises comparable within the published evaluation.
How the Scores Were Calculated
OpenAI checked design outputs in simulation. A design scored zero if it failed a check or used more than 16 times the original core’s area.
For passing designs, the company measured instructions per clock cycle on workloads hidden from the models. It took the geometric mean across those workloads, then multiplied it by a maximum clock frequency estimated using Yosys.
The resulting metric estimates instructions per second:
Estimated throughput = instructions per clock cycle × estimated clock frequency
This captures an important trade-off. A redesign could execute an instruction in fewer cycles while making each cycle longer. Counting cycles alone would miss that cost.
For each core, OpenAI normalized the best observed throughput under the same constraints to 1.0. The charts show average scores across runs, including failures that scored zero. A score of 1.0 therefore means the best observed reference in that evaluation, not a universal performance level or a percentage improvement over a production chip.
At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra had the highest average existing-core optimization score on both processors. OpenAI also reports that all five evaluated models scored higher when designing from requirements than when improving the existing cores. TNW’s coverage says GPT-6.2 Astra led both types of task.
Why Higher Scores Do Not Directly Predict Production Gains
The area limit is a major qualification. A design could pass while occupying up to 16 times the original core’s area. That makes the exercise useful for exploring throughput, but it gives models far more room than an area-saving production task would.
The published method also uses simulation checks and estimated clock frequency. Those are useful evaluation tools, but they answer narrower questions than checking all required behavior and confirming timing in a finished physical implementation.
A further caveat affects the apparent advantage of designing from scratch. TNW reports that OpenAI acknowledged the requirements-based task was easier because models could leave out features present in the originals. Keeping the test objectives constant does not necessarily require rebuilding every capability of the starting design.
My reading is that the benchmark rewards finding a design that performs well within the evaluator’s contract. Production engineering has a broader contract: required features, compatibility, implementation constraints and verification obligations must all survive.
The evaluation still shows useful model progress under stated conditions. A more production-representative test would tighten the area budget, require the full relevant feature set and carry candidates through more complete verification and physical implementation checks. That conclusion follows from the published conditions; it is not an independently reproduced result.
Synopsys Makes Verification Central to the Next Step
OpenAI says it is working with Synopsys to train GPT-Synopsys to use the electronic design automation tools engineers already rely on. The announcement describes training in progress, without announcing general availability.
The partnership fits the workflow behind the AMD results: propose an RTL change, check that it preserves behavior, and evaluate whether its implementation improves the desired trade-off. Better proposals matter, but reliable tool use is what lets engineers distinguish an attractive rewrite from a usable circuit.
Verification remains essential. TNW reports that OpenAI chip chief Richard Ho said the previous month that he still did not fully trust AI design changes and that every change was checked.
That caution is consistent with the production results. The valuable claim is that model-written changes can clear existing engineering checks and earn a place in real designs. Jalapeño’s bring-up and AMD’s manufacturing decision are the clearest evidence offered so far; the RISC-V scores explain capability progression under a separate, more permissive set of rules.
Frequently Asked Questions
3 questions
1
Has AMD Shipped Chips With These Changes?
OpenAI says some changes entered a design AMD sent for manufacturing. The announcement does not establish that those chips have shipped to customers.
2
Does the 14% Figure Mean a 14% Faster GPU?
No. It is a reported PPA improvement on flagship GPU IP, without enough detail to convert it into a whole-GPU speedup.
3
Is GPT-Synopsys Available to Use?
The announcement describes GPT-Synopsys as being trained with Synopsys. It does not announce general availability.