OpenAI Announces Up to 14% Gains on AMD's Flagship GPU

OpenAI claims to have achieved up to a 14% improvement in the trade-off between performance, power consumption, and area on AMD's flagship GPU, compared to already highly optimized benchmarks. Some modifications have been integrated into a design sent for fabrication. The company also presents evaluations on RISC-V processors, with reservations about their industrial representativeness.
Tests Allow for Fewer Functions
Designing a processor based on written requirements does not necessitate replicating all the functions of the original circuit. OpenAI emphasizes that its models can thus omit performance counters or additional instructions. These elements, which sometimes contribute little to the measured throughput, could explain part of the observed advantage. To align more closely with industrial constraints, the company deems it necessary to impose them, tighten the area limit, and control timing constraints after physical synthesis.
In this trial, the models have written requirements and examples of open-source code describing the interfaces with the system. They determine the data flow and the breakdown of operations into steps. Correctness checks, memory and cache parameters, as well as throughput objectives remain those of the trial consisting of modifying an existing circuit. According to OpenAI, around 112,000 tokens per execution, the five tested models achieve better scores with this method on both PicoRV32 and Ibex.
GPT-6.2 Astra benefits more on Ibex than on PicoRV32, the company indicates. Based on the PicoRV32 requirements, its average score reaches 0.7 near 112,000 tokens, then progresses in the next two budgets examined. For Ibex, it hovers around 0.75 at the same budget and remains superior to the other models as the common budgets increase.
AMD Retains Rewritten Circuits by Model
The partnership between AMD and OpenAI focuses on PPA, which encompasses performance, power consumption, and area. To improve these criteria on AMD hardware, the model rewrites the SystemVerilog RTL, a language describing circuits. OpenAI compares the progress made in less than 5 months to about a year of work. The company attributes it to expert microarchitectural reasoning and indicates that it can rewrite circuits for computation, control, or data transfer.
According to OpenAI, AMD has retained certain proposals for a design recently sent for fabrication and has included others in its production files for future versions. AMD has also verified percentage improvements in PPA in significant blocks of its upcoming flagship CPU, based on circuits written by the model. OpenAI anticipates further gains against the most optimized benchmark, without presenting them as already secured.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Synopsys tools were used by the model to check the behavioral equivalence of the modified blocks and to measure trade-offs between PPA criteria, OpenAI specifies. OpenAI and Synopsys are partnering to train GPT-Synopsys to use Synopsys's electronic design automation (EDA) tools like expert designers.
OpenAI Measures Throughput on Two Processors
To compare model generations, OpenAI selected two open-source 32-bit RISC-V processors capable of executing microcontroller-type tasks. PicoRV32 is a compact core intended for FPGAs, reprogrammable chips; Ibex is the embedded processor of the OpenTitan security chip. The entire evaluation totals 9,881 rated executions, divided between optimization and design tasks, cores, and various token budgets.
In the optimization trial, the model receives the original hardware description with the instruction to increase the instructions executed per second. This measure considers both the number of cycles required and their duration: an instruction requiring fewer cycles may be accompanied by longer cycles. OpenAI controls the designs through simulation and assigns a score of zero if they fail a verification or if their area exceeds 16 times that of the original core.
For accepted circuits, instructions per cycle are measured on tasks hidden from the models. Their geometric mean is multiplied by the maximum estimated frequency with Yosys to obtain an estimated throughput. The best throughput observed for each core, under identical constraints, receives a score of 1.0; the averages also include trials rated zero.
According to OpenAI, GPT-6.2 Astra leads on both cores around 112,000 tokens per trial, while GPT-6.1 Sol surpasses GPT-5.6 Sol on each. On the existing PicoRV32, GPT-6.2 Astra maintains the advantage at the highest comparable budgets and achieves, around 112,000 tokens, approximately twice the score of GPT-6 Astra. On the existing Ibex, its average result further progresses with the higher budgets tested. GPT-6 Astra and GPT-6.1 Sol are closer there.
Jalapeño Validated on Silicon
On Jalapeño, its first chip dedicated to inference, OpenAI indicates it began using GPT-4o in early 2025. The company announced in August that its models had accelerated development. According to the company, GPT-4o improved the PPA, and the commissioning confirmed the optimizations on silicon. The team estimates about a 13% reduction in area for the vector unit and about 10% for the matrix unit of standard cells.
The changes made to the vector unit, responsible for parallel computations, involve floating-point adders and multipliers, conversions, reduction logic, and the special function unit. According to OpenAI, the models proposed other logical organizations and different distributions of computations across pipeline stages. These explorations identified opportunities for hardware sharing and reduced intermediate storage.
In the matrix unit, the work concerns floating-point adders, multipliers, and circuits calculating low-precision dot products. OpenAI reports that a redesigned multiplier occupies about 15% less cell area at synthesis, with latency reduced from two clock cycles to one. A distinct adder saves about 15% area while maintaining its two-cycle latency. Additional savings come from modifications to dot product circuits and the arrangement of the pipeline.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.