OpenAI and AMD demonstrate concrete gains in chip design
OpenAI claims its models helped AMD improve the PPA of a flagship GPU by up to 14% in under five months. The company also accelerated Jalapeño and its tests on RISC-V cores, while preparing a tool with Synopsys.
OpenAI reports achieving up to a 14% gain in power, performance, and silicon area on high-end GPU blocks with AMD after less than five months of work. The company also states it advanced faster on Jalapeño, its first custom inference chip, and on processor design tests for RISC-V. This type of result matters because it addresses a very concrete bottleneck: the speed at which chips are designed, verified, and prepared for manufacturing.
Quand l’IA se met à écrire du RTL
The key point here lies in RTL SystemVerilog, the code describing chip circuits before manufacturing. OpenAI explains that its models proposed rewrites later verified with Synopsys tools to preserve block behavior while improving PPA, i.e., the trade-off between power consumption, performance, and area. On AMD's flagship GPU, some modifications were integrated into design files sent for fabrication.
The company speaks of a gain equivalent to “roughly a year of progress” achieved “in less than five months.” A useful phrase, but to be taken cautiously: OpenAI itself specifies that comparisons are based on already heavily optimized baselines. In other words, this is not about a prototype cobbled together in a garage but work on blocks already refined by very experienced engineering teams.
Jalapeño, le chip maison, a servi de terrain d’essai
OpenAI indicates it began using GPT-4o to help design Jalapeño, its first inference chip, in early 2025. In this project, the models reportedly reduced the standard cell area by about 13% in the warp engine, which manages and executes parallel tasks, and about 10% in the matrix unit, an arithmetic component. Explored approaches included reorganizing logic, distributing calculations across pipeline stages, and reducing intermediate storage.
In examples detailed by OpenAI, a multiplier in the matrix unit was redesigned to save about 15% area in synthesis while reducing latency from two clock cycles to one. Another redesign, on an adder, reduced area by about 15% without changing latency. Gains were then silicon-validated during Jalapeño's tuning, and OpenAI states that AI assistance shortened the time from initial design to tape-out to nine months, according to results published on August 25, 2026.
Des tests sur RISC-V pour comparer les modèles
To measure capability evolution, OpenAI also tested five generations of models on two open-source 32-bit RISC-V cores: PicoRV32, designed for FPGAs, and Ibex, used in OpenTitan. The evaluation covered both optimizing existing RTL and creating a core from written specifications and interface models. In total, 9,881 runs were scored across tasks, cores, and token budgets.
With a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra achieved the best average score on both cores for RTL optimization. GPT-6.1 Sol also outperformed GPT-5.6 Sol on both cores. In “specification to RTL” mode, GPT-6.2 Astra showed a normalized average score above 0.7 on PicoRV32 and about 0.75 on Ibex. All models performed better starting from specifications than optimizing existing RTL, at least in this test framework.
Les bons scores ne disent pas tout
OpenAI adds an important caveat: models starting from written specifications may omit features present in original cores, such as performance counters or certain instructions, if these elements do not significantly affect throughput measured on tested workloads. In this context, part of the advantage may come from these omissions. To approach industrial reality, these functions would need to be enforced, area limits tightened, and timing verified after physical synthesis.
The subject then becomes less spectacular than it seems but more interesting for industry. If AI can propose valid variants, the whole challenge is to integrate them into design flows where verification remains non-negotiable. It is not the time to let a chip go to fabrication with a nice intuition but a bad calculation.
Synopsys veut brancher les modèles sur les vrais outils EDA
OpenAI is preparing, with Synopsys, a model called GPT-Synopsys, intended to use electronic design automation (EDA) tools as specialized engineers do. The joint development agreement, announced on September 30, 2026, also covers marketing and revenue sharing. Synopsys states its engineers will be able to delegate objectives ranging from PPA optimization to timing closure and verification, while agents run tools, interpret results, apply changes, and iterate until a verified outcome.
Synopsys specifies that customer data will not be used to train GPT-Synopsys, with encryption at rest and in transit, as well as configurable retention, audit, and permission controls. At this stage, this is the most serious missing piece: linking observed benchmark gains to industrial workflows without breaking verification safeguards.
Cette course ne concerne pas seulement OpenAI
The movement goes beyond OpenAI and AMD. According to the source, Anthropic also uses AI in chip design: its Mythos model reportedly processed prompt engineering for AMD MI355X GPUs in three days, compared to six to twelve months usually, according to the company. It also runs Claude on four chip families to optimize cost and performance per dollar. The market thus sees less a gadget than the next layer of automation for intensive computing, where every month saved can change a product roadmap.
One simple question remains: if models can already propose useful redesigns, how far will they go before becoming a standard chip design tool? The gains described here show at least one thing: in semiconductors, AI no longer just helps write code; it is starting to touch the silicon itself.
Key points
- Up to 14% PPA gain on AMD GPUs, according to OpenAI.
- Jalapeño reached tape-out in nine months, according to OpenAI.
- 9,881 runs were used for OpenAI's RISC-V tests.
- GPT-6.2 Astra dominates tested models on PicoRV32 and Ibex.
- Synopsys and OpenAI announced GPT-Synopsys on September 30, 2026.
- Anthropic cites three days versus six to twelve months for Mythos.
By the numbers
- 14% — maximum PPA gain announced on AMD GPU, November 2026, OpenAI.
- 13% — estimated area reduction in Jalapeño's warp engine.
- 9,881 — runs scored in RISC-V evaluation.
- 112,000 tokens — comparable budget per run used in tests.
- September 30, 2026 — announcement of OpenAI-Synopsys agreement.