Chips & Compute

OpenAI says its models cut AMD flagship GPU IP power-performance gaps by 14%

OpenAI says its models delivered up to 14% PPA gains on AMD flagship GPU IP in under 5 months, and 15% area cuts on its own Jalapeño inference chip blocks.

OpenAI Details AI’s Role In Chip Design
OpenAI Details AI’s Role In Chip DesignKlemens Morbe / unsplash
By James Calloway5 min read

Why it matters

  • OpenAI reports up to 14% PPA improvement on AMD's flagship GPU IP, roughly a year of progress in under 5 months
  • GPT-4o, used since early 2025, produced ~15% cell-area reductions in multiplier and adder redesigns on the Jalapeño chip
  • AMD has already included some AI-suggested changes in a design sent for manufacturing
  • GPT-6.2 Astra scored highest on RISC-V core tasks at ~112,000 tokens per run across 9,881 scored evaluations
  • All five models tested scored higher designing RISC-V cores from specifications than optimizing existing PicoRV32 and Ibex designs

OpenAI has published detailed results showing its models achieved up to 14% improvement on AMD's flagship GPU IP relative to heavily optimized baselines — work AMD has already folded into a design sent for manufacturing.

The company pulled back the curtain in a post titled "Advancing Chip Design With Our Models," describing how AI is shaping Jalapeño, OpenAI's first custom inference chip (a chip built to run trained models rather than train them). The disclosures arrive as chip design costs balloon and AI companies race to control their own silicon. If the numbers hold, models could compress design cycles that traditionally take engineering teams a year or more.

What did GPT-4o do on the Jalapeño chip?

OpenAI says it began using GPT-4o in early 2025 to help design Jalapeño. The gains clustered in two blocks of the chip.

The vector unit — which manages and executes parallel computing tasks — saw savings from floating-point adders and multipliers, conversion functions, reduction logic, and the special-function unit. "The models explored both alternative logic implementations and how to divide computations across pipeline stages, finding opportunities to share hardware and reduce intermediate storage," OpenAI wrote.

The matrix unit, an arithmetic component, delivered more measurable wins. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while keeping its two-cycle latency.

Those are concrete, physical savings: less silicon area and fewer clock cycles translate directly into cost and speed for inference hardware. The takeaway is that OpenAI's own chip is already being shaped by its own models.

How big are the AMD gains?

The AMD partnership produced the headline number: up to 14% PPA improvement — PPA meaning power, performance, and area, the three axes chip engineers optimize — on AMD's flagship GPU IP.

OpenAI attributes part of the gain to its model rewriting SystemVerilog RTL, the code that describes a chip's circuits. "This represents roughly a year of progress achieved in less than 5 months–unprecedented results at the frontier of chip design," the company said.

That claim carries weight because AMD acted on it. OpenAI states AMD has already included some of these changes in a recent design it sent for manufacturing, and added others to its production design files for future versions. A partner committing changes to tape-out (the stage where a design is finalized for fabrication) is stronger evidence than a benchmark alone.

The caveat: all figures come from OpenAI's own account. Neither the 14% figure nor the "year of progress" framing has been independently verified, and OpenAI did not specify which flagship GPU IP benefited. The takeaway is that AMD is shipping AI-suggested changes, which is the hardest evidence in the post.

What happened in the RISC-V experiments?

OpenAI also tested its models on two existing open-source RISC-V cores (RISC-V is an open instruction-set architecture; a "core" is the basic processor engine) — PicoRV32 and Ibex — plus a from-scratch design task.

The evaluation method was strict, at least on paper:

  • Designs scored zero if they failed a correctness check or used more than 16 times the original core's area.
  • Passing designs were measured on instructions per clock cycle across workloads hidden from the models.
  • OpenAI took the geometric mean across workloads and multiplied it by the maximum clock frequency estimated with Yosys, an open-source synthesis tool, to estimate instructions per second.
  • Each core's best observed throughput under the same constraints received a score of 1.0.

With a budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both cores. GPT-6.1 Sol also outscored GPT-5.6 Sol on both. The full evaluation covered 9,881 scored runs across both tasks, both cores, and all token budgets.

Nine thousand-plus runs is a serious sample, but the scoring pipeline still relies on OpenAI's own simulation harness and Yosys-based frequency estimates rather than fabricated silicon. The takeaway: newer models beat older ones on real hardware design tasks, at least in simulation.

Did designing from scratch beat optimizing existing chips?

Yes, and OpenAI flags this as the most telling result. When models designed a RISC-V core from a written specification rather than improving existing code, the gains were larger.

For that task, OpenAI gave each model written requirements and templates from the open-source code showing how the core connects to the rest of the system. The model then chose how data would move through the core and how to divide the work into pipeline stages. Correctness checks, memory and cache settings, and throughput objectives stayed identical to the optimization task.

At roughly 112,000 tokens per run, all five models tested scored higher when designing from written requirements than when improving existing designs — and that held for both PicoRV32 and Ibex. GPT-6.2 Astra gained more on Ibex than on PicoRV32.

The result suggests models are less constrained when they are not anchored to human-written code, which hints at where AI-driven design goes next: greenfield design rather than incremental patching. The counterweight is that these are simulated scores against specification targets, not fabricated, validated chips.

Why the stakes extend past OpenAI

The economics matter. Custom silicon has historically required large engineering teams and multi-year cycles. Google has used AI-assisted methods in its TPU (Tensor Processing Unit) design flow for years; OpenAI's post shows the technique spreading to frontier model developers and their partners. A 14% PPA gain on already-heavily-optimized GPU IP — if independently confirmed — would force every major chip vendor to reassess its design toolchain.

For OpenAI specifically, the work supports its vertical integration push: Jalapeño is its first custom inference chip, and AMD is a key supplier beyond Nvidia. Using its own models to squeeze more out of both its silicon and its partner's changes the cost structure of running frontier AI at scale.

What is not said is equally notable. OpenAI does not disclose absolute scores, absolute areas, or instructions-per-second figures for the RISC-V cores; it reports normalized scores only. It does not name the AMD flagship IP block. And it does not say when Jalapeño tapes out or reaches production.

The clearest signal in the post is AMD shipping AI-written RTL changes to a fab — that is design automation crossing from research claim into production silicon.

Source: webpronews.com

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. OpenAI models deliver up to 14% AMD PPA gain and Jalapeño area cuts
  2. OpenAI Reports Up to 14% PPA Gains in AMD Chip Designs
  3. OpenAI's Jalapeño Chip: LLMs Cut Design Time to Record Lows
  4. OpenAI Posts First Benchmark Results for Jalapeño Custom Inference Chip
  5. OpenAI and Broadcom Unveil Jalapeño, a Custom LLM Inference Chip

« Previous articleNext article »