Ai Engineering 7 min read

OpenAI's models rewrote chip RTL for AMD and its own Jalapeño accelerator: up to 14% PPA, some of it already taped out

OpenAI says its models rewrote SystemVerilog for AMD GPU and CPU blocks and for its Jalapeño accelerator, with gains of up to 14% on AMD GPU IP and 10-15% area cuts on Jalapeño blocks. What was measured, what was verified, and what is still a claim.

OpenAI published a technical report this week, Advancing chip design with our models, dated November 18 in the earliest coverage. It says its language models have been rewriting the register-transfer level (RTL) code of real chips, not toy benchmarks: blocks of AMD’s flagship GPU IP, key blocks of an upcoming AMD CPU, and OpenAI’s own Jalapeño accelerator. The headline figure is up to 14% better aggregate power, performance and area (PPA) on AMD GPU blocks, measured against designs that were already heavily optimized. AMD has put some of the model-generated changes into a design that has gone to manufacturing.

If it holds up, this matters because RTL optimization is one of the most expert-limited steps in building a chip. OpenAI’s own estimate is that the work delivered roughly a year of design progress in less than five months. It is also a claim from a company with a financial stake in AMD’s success, so it is worth separating what was measured from what was asserted.

What OpenAI says the models did

The models worked directly on SystemVerilog, the hardware description language most chip blocks are written in. They were given existing RTL and asked to make it smaller, faster or less power hungry without changing what the block computes. According to Inside AI’s summary of the report, the changes were the kind a senior microarchitect makes: reworking adders, multipliers, number-format conversion functions and reduction logic, splitting computation differently across pipeline stages, sharing hardware between operations and cutting intermediate storage. OpenAI describes watching a model “reason about RTL at the level of an expert microarchitect.”

The reported results, by project:

  • Jalapeño (OpenAI’s own accelerator): about 13% less standard-cell area in one block (reported as the vector unit) and about 10% in the matrix unit. Inside AI and StartupHub both say this early work used GPT-4o; one other outlet names a different model, so treat the model name as unconfirmed until you read the report itself. Inside AI says the gains were proven in silicon during Jalapeño’s bring-up.
  • Individual arithmetic units: one multiplier redesign used about 15% less cell area in synthesis and cut latency from two clock cycles to one. A separate adder redesign also saved about 15% of cell area at the same two-cycle latency.
  • AMD GPU IP: up to 14% aggregate PPA improvement on blocks of AMD’s flagship GPU IP. AMD included some changes in a recent design sent for manufacturing and added others to production design files for future versions.
  • AMD CPU: AMD verified single-digit percentage PPA gains across key blocks of an upcoming flagship CPU.

Every rewrite went through Synopsys EDA tools, which checked that the block’s behavior was preserved and measured the PPA trade-off. OpenAI says it is now working with Synopsys on GPT-Synopsys, a model trained to drive Synopsys tools the way an experienced designer does.

The benchmark behind the claims

The report also includes a controlled evaluation that is easier to judge than the production results, because its rules are public. OpenAI took two open RISC-V cores, PicoRV32 and lowRISC’s Ibex from the OpenTitan project, gave each model the original hardware description, and asked it to raise throughput measured in instructions per second. A design scored zero if it failed simulation checks or used more than 16 times the original core’s area. Runs used about 112,000 tokens each, and OpenAI scored 9,881 of them.

GPT-6.2 Astra led on both cores, and GPT-6.1 Sol outscored GPT-5.6 Sol, which fits the broader pattern of newer models doing better at long, verifiable engineering tasks. In a second task, models designed a core from written requirements and templates, and all five tested models scored higher designing from scratch than when optimizing the existing RTL.

OpenAI is clear about the limits of that test. It let models drop features such as performance counters, and a production-grade version would need mandatory features, tighter area limits and timing checks after physical synthesis. A 16x area budget is generous; real silicon teams fight over single-digit percentages.

Why chip RTL suits language models

Chip design has a property most software lacks: an unforgiving, automated referee. Equivalence checking, simulation and synthesis tell you whether a rewritten block still computes the same function and exactly how much area, timing and power it costs. That is the same structure that has made coding agents useful: a model proposes, a deterministic tool scores, and the loop repeats. A model does not need to be right the first time if a verifier catches its mistakes cheaply.

PPA work is also a search problem across a huge space of equivalent circuits. Human teams explore a small part of it because each idea takes days of an expensive engineer’s time. A model that can try hundreds of microarchitectural variants overnight, with the EDA flow rejecting the broken ones, changes the economics even if most of its proposals are bad. The 15% multiplier result is a good example of where the gains come from: arithmetic units are replicated thousands of times across a GPU, so a small cell-area saving per unit compounds across the die.

The hard parts of chip design are mostly untouched by this report. Physical design, timing closure across a full chip, power integrity and verification of whole subsystems are where schedules slip, and OpenAI’s results are block-level. The report says the approach still depends on human oversight and robust verification flows.

How much weight to put on 14%

Three caveats come from the coverage itself. First, “up to 14%” is a ceiling on specific GPU blocks, and Yahoo Finance’s analysis notes OpenAI did not say how much of it is power, performance or area. Second, nobody has reported measured gains on finished AMD silicon; a taped-out design is still a long way from a shipping product. Third, the parties are not neutral. Under their October 2025 agreement, OpenAI committed to deploying six gigawatts of AMD GPUs and received warrants for up to 160 million AMD shares that vest partly on AMD’s stock price. AMD has not issued a formal statement; CEO Lisa Su posted publicly, quoting Greg Brockman’s note about the taped-out design and praised the engineers at both companies.

None of that makes the results wrong. The Jalapeño figures are on OpenAI’s own chip, the Synopsys checks are standard industry tooling, and AMD putting changes into a manufacturing design is a stronger signal than any press release. But it is a vendor report, not an independent study.

What to watch

The first test is whether AMD names a product. AMD presents at investor conferences in early December, and a named chip with a ship date carrying model-generated RTL would turn this from a claim into a fact.

The second is GPT-Synopsys. If OpenAI’s chip work becomes a feature inside the EDA tools most of the industry already uses, AMD’s head start shrinks to months and every chip team gets the same capability. Cadence and Siemens will be under pressure to answer.

The third is independent replication. The RISC-V benchmark is described in enough detail that academic groups can rerun it with open models on the same cores. If those results land near OpenAI’s, the case that frontier models can do real microarchitecture work gets much harder to dismiss.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

How Function Calling Works in LLMs

Function calling lets LLMs interact with external systems by requesting structured tool executions. Here's how the loop works, how to define tools, and what to watch for across providers.

Ai Engineering

GPT-6.2 Astra and Sol launch: 20% cheaper Astra, and OpenAI's first public model rated High for AI self-improvement

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, two months after shelving GPT-6.1 Astra. Astra drops to $8/$40 per million tokens, claims new highs on professional-work benchmarks, and is the first model OpenAI has shipped at its High threshold for AI self-improvement.

Ai Engineering

OpenAI folds its robotics team into research after GPT-6 Astra starts driving robots directly

Fortune reported on November 13 that OpenAI has disbanded its stand-alone robotics team and moved its researchers under Mark Chen and Greg Brockman, on the argument that general frontier models can now control robots without robot-specific training. Here is the evidence, the history and the counterarguments.

Ai Engineering

OpenAI Claims Its Model Falsified the Exponential Time Hypothesis. The Verifying Starts Now.

OpenAI announced October 26 that an internal reasoning model produced a Lean-checked 2^o(n)-time algorithm for 3-SAT, which would falsify the Exponential Time Hypothesis, alongside claimed Yang-Mills and Kakeya results, with academic shepherding as the verification mechanism and a prior 722-paper release that saw three withdrawals as the reason for skepticism.

Ai Engineering

The Data Behind the LLM Price War: OpenAI and Anthropic Cut Costs Up to 80% in Three Months

Max Woolf's October 25 analysis traces every major LLM price cut from June through September: the GPT-6.1 Sol family ended roughly 75% cheaper cumulatively, cache-read pricing drove most of it, and the Vercel gateway data shows costs falling to 43% of April levels while spend climbs.