OpenAI's Models Are Now Editing AMD's Flagship GPU RTL: What the 14% PPA Claim Does and Doesn't Show

On 18 November 2026 OpenAI published “Advancing chip design with OpenAI models”, a short technical post with a large claim: its models are already changing the hardware description code of production chips, not just demo circuits. Three results are reported:

  1. AMD flagship GPU IP: up to 14% improvement in power, performance and area (PPA) “relative to heavily optimized baselines”, produced by the model rewriting SystemVerilog RTL. OpenAI says AMD “has already included some of these changes in a recent design it sent for manufacturing” and added others to production design files for future versions. AMD also “verified single-digit percentage gains in PPA across key blocks” of its upcoming next-generation flagship CPU using model-written designs.
  2. Jalapeño, OpenAI’s own inference chip: model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and OpenAI says those optimizations were “silicon proven” during bring-up.
  3. A public benchmark: five OpenAI model generations were scored on optimizing and re-designing two open-source RISC-V cores, PicoRV32 and Ibex, over 9,881 scored runs. GPT‑6.2 Astra led.

Everything above is OpenAI’s own account. AMD had not published a separate statement when The Next Web and AI Weekly covered the post; both relay OpenAI’s numbers rather than independent measurements. Keep that in mind for the rest of this article.

What the models actually did

PPA is the three-way trade-off every chip block lives under: how much power it draws, how fast it runs, how much silicon it occupies. RTL (register-transfer level) code, in SystemVerilog or Verilog, describes the circuit at the level of registers, adders, multiplexers and pipeline stages. Synthesis tools turn it into standard cells, and that cell count is the “area” in the numbers.

The concrete examples OpenAI gives from Jalapeño are the sort of work a senior microarchitect does by hand:

  • In the vector unit, savings came from floating-point adders and multipliers, format-conversion logic, reduction trees and the special-function unit. The models tried alternative logic implementations and re-split computations across pipeline stages, finding places to share hardware and reduce intermediate storage (fewer pipeline registers).
  • In the matrix unit, one multiplier redesign used about 15% less cell area while cutting latency from two clock cycles to one. Smaller and faster is the unusual part, since designers normally trade one for the other. A separate adder redesign saved about 15% area at the same two-cycle latency.

Notably, OpenAI says this started with GPT‑4o in early 2025, a model now two generations old. In its August post on Jalapeño’s first results, the company said AI help took the chip from initial design to tapeout in nine months.

For AMD, the loop is described in one sentence that matters more than the headline number: the model “used Synopsys tools to verify that each rewrite preserved the block’s behavior and to evaluate the trade-offs in power, performance, and area.” That is the familiar agent pattern developers know from coding tools: the model proposes a diff, deterministic tools (equivalence checking and synthesis here, instead of unit tests and a compiler) decide whether it stands. The model’s creativity is only useful because the verifier is trustworthy.

The RISC-V benchmark, and the caveat OpenAI wrote itself

Production results at AMD can’t be reproduced by outsiders, so OpenAI also built a measurable proxy on two open-source 32-bit cores: PicoRV32 (a compact FPGA-oriented core) and Ibex (the embedded core used in the OpenTitan security chip).

  • Task A, optimize existing RTL: raise throughput, meaning instructions per second. A design scored zero if it failed a simulation check or used more than 16× the original area. Passing designs were scored as geometric-mean IPC on hidden workloads × maximum clock frequency estimated with Yosys, normalized so the best observed design for each core is 1.0. Averages include the zeros.
  • Task B, design from a specification: the model gets written requirements plus interface templates and chooses its own microarchitecture and pipeline staging.

At a comparable budget of roughly 112,000 tokens per run, GPT‑6.2 Astra had the highest average on both cores. On PicoRV32 its optimization score was about twice GPT‑6 Astra’s. Designing from spec, it averaged above 0.7 on PicoRV32 and about 0.75 on Ibex. GPT‑6.1 Sol beat GPT‑5.6 Sol on both cores, so the trend across generations is consistent.

The most interesting result is the one that cuts against the hype: all five models scored higher when designing from a spec than when improving the existing design. OpenAI is candid about why: starting from requirements lets models drop features such as performance counters and extra instructions that cost area but don’t help on the test workloads. It adds that requiring feature parity, tightening the area limit and checking timing after physical synthesis “would make the test more representative of production chip design.” In other words, the public benchmark is a measure of progress across model generations, not a stand-in for sign-off. A 16× area ceiling is generous, and a Yosys frequency estimate is not post-layout timing closure.

Why this matters beyond chip companies

AI is starting to improve the hardware it runs on. Jalapeño is an inference chip; AMD’s GPUs are what OpenAI and others train and serve on. If models can squeeze an extra 10–14% PPA out of blocks that expert teams had already optimized, that is cheaper compute for the next model, a feedback loop the AI-safety literature has long discussed in the abstract. Here it is a line item.

The pattern is the same one software teams are adopting. “Model proposes, verifier disposes” is how AI coding agents are being made safe enough to merge: tests, type checkers, linters, CI. Hardware happens to have unusually strong verifiers (formal equivalence checking) and an unusually expensive failure mode (a mask set), which may be why it’s a good early domain. The lesson for software is that the value of an agent is bounded by the quality of the checks around it.

It’s becoming a product. The post closes by pointing to GPT‑Synopsys, announced with Synopsys on 30 September 2026: a specialized model trained to operate Synopsys EDA tools “the way expert chip designers do”, under a multi-year deal with revenue sharing. Synopsys says customer design data won’t be used for training and will be encrypted at rest and in transit. The 18 November post is effectively the evidence OpenAI needs to sell that service.

Open questions

  • No AMD-side numbers. The post doesn’t name the GPU or CPU products, the specific blocks, or which of the 14% is power versus performance versus area. “Up to 14%” is a best case. And “sent for manufacturing” isn’t measured silicon; only the Jalapeño gains are described as silicon-proven.
  • Which model? The RISC-V charts name GPT‑6.2 Astra, but the post doesn’t say which model or internal tooling did the AMD work, or how much human expert steering was involved.
  • Hard to replicate. The RISC-V harness, hidden workloads and prompts aren’t published, so outsiders can’t rerun the 9,881 runs.
  • Trust is still manual. As The Next Web points out, OpenAI’s own hardware lead Richard Ho said last month that he still doesn’t fully trust AI-made design changes, so every one is checked. That’s healthy, and it also means the “year of progress in five months” figure includes a lot of human review time that the post doesn’t measure.

The fair reading: this isn’t AI designing chips on its own. It’s frontier models acting as tireless micro-optimizers inside a mature, tool-verified flow, finding gains experts left on the table. That’s less cinematic and more useful, and it’s the same shape the next few years of AI-assisted software engineering are likely to take.

Sources