OpenAIโs latest research demonstrates that large language models can actively participate in the intricate process of semiconductor creation, turning a field once dominated by seasoned engineers into a collaborative effort between humans and algorithms. In August 2026 the organization disclosed how its frontier models accelerated the development of a custom inference processor named Jalapeรฑo, a chip built specifically to serve AI workloads with high efficiency. By feeding the model detailed specifications of the target architecture, engineers were able to let the system suggest refinements that shaved down power consumption, boosted performance, and trimmed silicon area without sacrificing functionality. This breakthrough signals a paradigm shift: instead of relying solely on human intuition and years of trialโandโerror, AI can now generate concrete circuitโlevel ideas, test them through simulation, and propose revisions that are ready for tapeโout. The implication is profound for the entire semiconductor ecosystem, as it promises faster design cycles, lower engineering costs, and the possibility of exploring architectures that might have been deemed too risky or timeโconsuming just a few years ago. These advances come at a time when the industry faces mounting pressure to deliver more compute per watt while navigating the physical limits of lithography and the rising costs of advanced process nodes. By integrating AI into the design flow, companies can tap into a knowledge base that spans countless design iterations, enabling rapid exploration of tradeโoffs that would otherwise require extensive manual effort.
Early in 2025 OpenAI began a focused partnership with Advanced Micro Devices, providing its models to assist in the optimization of AMDโs flagship GPU intellectual property. The goal was to push the powerโperformanceโarea triangle toward more favorable values, a metric that directly influences silicon yield and market competitiveness. Using an early version of the model, engineers observed that even modest modifications to the RTL description could yield measurable gains in clock frequency while keeping power budgets steady. Over the course of several months the model identified opportunities to share hardware resources across pipeline stages, to replace bulky combinational blocks with more compact alternatives, and to restructure data paths so that fewer registers were needed. These changes translated into a reduction of standardโcell area in critical blocks and a measurable drop in dynamic power. The collaboration demonstrated that AIโdriven suggestions are not merely theoretical; they can be incorporated into productionโready design flows, reviewed with existing electronic design automation tools, and validated during silicon bringโup. This early success laid the groundwork for the deeper investigations reported later in the year, where the same models were applied to a broader set of processor components and to thirdโparty hardware. The partnership also opened the door to joint research on nextโgeneration compute units, positioning both companies to coโdevelop chips that could meet the demanding requirements of emerging AI services.
Within the Jalapeรฑo architecture the vector unit serves as the engine that executes parallel arithmetic operations, making it a natural target for AIโguided optimization. The model examined each component of the unit, from the floatingโpoint adder and multiplier pipelines to the conversion routines that translate between formats and the reduction logic that aggregates partial results. By exploring alternative logic families and by reโpartitioning computations across pipeline stages, the system discovered ways to reuse existing hardware rather than instantiating dedicated blocks for each operation. For example, it suggested merging the floatingโpoint adder with a shared carry chain, thereby cutting the number of distinct transistor devices required. The model also proposed dividing longโlatency multiplyโaccumulate operations into smaller, overlapping subโoperations, which reduced the need for large intermediate storage elements. These refinements collectively trimmed the standardโcell area of the vector unit by roughly thirteen percent, a gain that was later confirmed on silicon during the bringโup phase. In addition to area savings, the redesigns lowered the average power per operation, delivering a more energyโefficient core without compromising throughput. The efficiency gains also extended to the specialโfunction unit, where the model identified opportunities to fuse transcendental calculations with nearby arithmetic blocks, reducing instruction latency and freeing pipeline slots for other tasks.
The matrix unit, responsible for dense linearโalgebra operations that underpin most AI inference workloads, presented another fertile ground for modelโdriven redesign. By analyzing the floatingโpoint adder and multiplier structures, the system proposed a multiplier layout that required roughly fifteen percent less cell area after synthesis while halving the latency from two clock cycles to a single cycle. A parallel adder revision achieved a similar fifteen percent reduction in area while preserving the original twoโcycle latency, demonstrating that size and speed are not mutually exclusive. Moreover, the model reโengineered the lowโprecision dotโproduct circuitry, consolidating several small adders into a unified structure and reโordering pipeline stages to minimize data movement. These adjustments collectively shaved ten percent off the matrix unitโs standardโcell footprint and further improved its energy efficiency. The ripple effect was evident across the entire chip, as smaller arithmetic blocks freed up routing resources for other functional units, contributing to a more balanced overall design. Additionally, the model suggested reโbalancing the operand width of the dotโproduct unit to match the precision requirements of the targeted AI models, which eliminated unnecessary bits and further reduced power consumption without affecting numerical accuracy. These enhancements not only lowered the silicon area but also contributed to a measurable decrease in the chipโs overall dynamic power, which is critical for maintaining acceptable thermal margins in highโthroughput AI servers.
Overall, the AIโassisted design effort delivered quantifiable improvements across the Jalapeรฑo compute architecture. The vector unit saw a thirteen percent reduction in standardโcell area, while the matrix unit achieved a ten percent shrinkage, translating into a noticeable contraction of the chipโs total die size. These percentages were derived from postโsynthesis reports that compared the AIโproposed layouts against the baseline handโcrafted versions. The silicon bringโup phase confirmed that the revised blocks behaved as expected, with no functional regressions observed during extensive testโpattern execution. Power measurements indicated a proportional drop in dynamic energy per operation, and timing analyses showed that the critical paths remained within the original performance envelope. Such validation is essential because theoretical area savings alone do not guarantee realโworld benefits; the chip must meet its timing, power, and functionality targets in actual silicon. The successful tapeโout of Jalapeรฑo thus serves as a concrete proof point that large language models can contribute meaningfully to productionโgrade chip design, moving beyond speculative experiments toward deployable results. The design team also reported that the accelerated development timeline, which shrank what would have been a yearโlong optimization cycle into just five months, allowed the company to allocate resources more flexibly and respond faster to market demands for higher efficiency AI accelerators.
Building on the Jalapeรฑo experience, OpenAI entered a deeper collaboration with AMD to apply its models to the companyโs flagship GPU intellectual property. By feeding the latest model variants the detailed RTL descriptions of the GPUโs compute clusters, the team observed up to fourteen percent improvement in the powerโperformanceโarea metrics relative to previously heavily tuned baselines. The breakthrough came from rewriting portions of the SystemVerilog code that define the circuit topology, a task that traditionally requires many months of expert microarchitect effort. In less than five months the model generated alternative netlists that preserved functional behavior while reducing both transistor count and clockโtoโclock latency. AMD incorporated several of these modifications into a design that has already progressed to the manufacturing stage, and it added additional changes to its production library for future silicon revisions. The partnership also enabled AMD to verify singleโdigit percentage gains in PPA across key blocks of its upcoming nextโgeneration flagship CPU, demonstrating that the same methodology can be transferred across distinct product families. These results suggest that AI can act as a catalyst for accelerating IP development cycles, reducing timeโtoโmarket and enabling smaller design teams to compete with larger vendors.
The model demonstrated an impressive capacity to reason about RTL at a level that rivals seasoned microarchitects, enabling it to rewrite circuits that execute arithmetic, control flow, or data movement with confidence. By leveraging Synopsys electronic design automation tools, the system automatically verified that each proposed alteration preserved the functional equivalence of the original block, while also quantifying the impact on power, performance, and area. This dualโobjective verification allowed engineers to explore aggressive optimizations without fearing hidden bugs, a capability that historically required extensive manual simulation and formal checking. The modelโs suggestions often involved subtle changes such as reโbalancing carry chains, reโordering pipeline stages, or consolidating adjacent logic cells, all of which were validated through synthesis and timing reports. Moreover, the ability to evaluate tradeโoffs in a systematic way helped the design team prioritize modifications that delivered the highest return on investment, ensuring that the final tapeโout met stringent commercial targets. This systematic approach also facilitated the generation of comprehensive regression suites, enabling continuous verification as the design evolved and reducing the risk of lateโstage silicon errata that can be costly to resolve.
AMDโs validation of the modelโgenerated designs extended beyond the GPU, reaching the upcoming flagship CPU architecture. Independent testing showed that key blocks of the new processor exhibited singleโdigit percentage improvements in power, performance, and area, confirming that the AIโcrafted RTL modifications translate across different compute units. These gains, while modest in isolation, collectively contribute to a more efficient silicon footprint and higher clockโfrequency headroom, both of which are critical for meeting the performance expectations of nextโgeneration dataโcenter and edge workloads. Looking ahead, the collaboration anticipates further refinements that could push PPA improvements beyond the current singleโdigit range, especially as the model continues to mature and gain exposure to more sophisticated design constraints. The expectation is that future iterations will enable tighter integration between AI suggestions and the advanced layoutโaware tools used by semiconductor foundries, thereby unlocking even greater optimization opportunities. This trajectory positions AIโassisted design as a longโterm enabler for the semiconductor industry, helping companies stay ahead of the curve in an era where each new node brings diminishing returns on raw transistor scaling. Such forwardโlooking improvements could also reduce the overall cost of ownership for customers, as lower power consumption and higher performance translate into smaller cooling solutions and longer device lifespans, reinforcing the economic incentive for AIโdriven design innovation.
To assess how generative models have progressed in handling complete processor architectures, OpenAI selected two openโsource 32โbit RISCโV cores for a systematic evaluation: PicoRV32, a compact core intended for FPGAโbased prototyping, and Ibex, the embedded processor employed in the OpenTitan security chip. Each core was presented to the models in its original hardware description language, and the task was to increase its instruction throughputโthe number of instructions executed per secondโwhile respecting strict constraints on area and correctness. The evaluation budget limited each model run to roughly 112,000 tokens, a figure chosen to reflect the amount of context a model can comfortably process in a single pass. Designs that failed simulation checks or exceeded sixteen times the original coreโs area received a score of zero, ensuring that only viable architectures were considered. For valid designs, the team measured instructions per clock cycle on a set of hidden workloads, combined this metric with an estimated maximum clock frequency derived from Yosys, and then calculated an overall instructionsโperโsecond score. A perfect score of one represented the best achievable throughput under the given constraints, allowing for direct comparison across model versions and core types. The methodology also incorporated a geometric mean across the workload set to smooth out variability, and the token budget was deliberately set to challenge the models without overwhelming them, thereby revealing their true capability limits.
Across the evaluation, GPTโ6.2 Astra consistently achieved the highest average scores on both PicoRV32 and Ibex, especially when operating with the larger token budget of approximately 112,000 tokens per run. On PicoRV32, its mean score reached roughly twice that of the earlier GPTโ6 Astra model at comparable token limits, indicating substantial improvement in generating efficient pipeline optimizations. For Ibex, GPTโ6.2 Astraโs score continued to climb as the token budget increased, showing a clear upward trend that outperformed GPTโ6.1 Sol and GPTโ5.6 Sol at the same budgets. The charts presented in the research illustrate that the modelโs ability to reason about architectural tradeโoffs grew with each generation, allowing it to propose more aggressive yet still viable microarchitectural changes. These results underscore that the latest models are not only better at reโusing existing structures but also at inventing novel configurations that achieve higher throughput without breaching the area or correctness limits imposed by the evaluation framework. The performance gap between the topโscoring model and its predecessors grew by more than thirty percent, highlighting the rapid advancement in the modelsโ architectural reasoning capabilities and their increasing suitability for realโworld chip design tasks. This progress suggests that AI can soon become a standard tool in the semiconductor design ecosystem.
In the second phase of the assessment, each model was given highโlevel requirements and code templates that described how a core should interface with the surrounding system, rather than being handed the complete RTL of an existing core. This approach allowed the models to decide which architectural features to retain and which to discard, such as performance counters or rarely used instructions that have minimal impact on the selected workloads. By omitting these legacy elements, the models could concentrate on the core functional units that directly affect instruction throughput, often resulting in more compact and efficient designs. Consequently, all five models achieved higher scores when constructing a core from the written specifications compared to attempting to improve the preโexisting RTL, with GPTโ6.2 Astra showing the most pronounced gain on the Ibex core. The 9,881 scored runs across both tasks and both cores demonstrated that the freedom to redesign rather than merely tweak existing structures unlocks greater design space, leading to more innovative and higherโperforming architectures. These findings imply that designers should consider providing flexible specification inputs to AI assistants, as this can dramatically expand the range of viable solutions and accelerate the discovery of highโefficiency microarchitectures. In practice, this approach shortens development cycles and reduces the need for extensive manual optimization effort.
For engineers and organizations seeking to harness AI in chip design, the key takeaway is to integrate generative models early in the design flow, treating them as collaborative partners rather than mere novelty tools. Begin by defining clear performance targets and providing the model with highโquality RTL or specification documents that capture the essential functional and geometric constraints. Use established EDA platforms, such as those from Synopsys, to run automated verification after each AIโgenerated modification, ensuring that power, performance, and area goals remain met. Adopt a phased approach: start with small, wellโbounded blocks like the vector or matrix units, validate the modelโs suggestions, and then scale the technique to larger subsystems such as entire CPU cores or GPU clusters. Finally, monitor the modelโs evolution, as newer generations promise even tighter integration with layoutโaware tools and more accurate tradeโoff analysis, making AI an increasingly indispensable component of modern semiconductor development pipelines. By embedding AI assistance into existing design reviews and continuous integration pipelines, teams can achieve faster iteration cycles, reduce manual debugging effort, and ultimately deliver higherโperformance chips to market with lower development costs. Adopting these practices now positions companies to stay competitive as the semiconductor industry embraces AIโdriven innovation and faces tighter timelines for new product introductions for sustained longโterm growth.