Chips & Compute

OpenAI models deliver up to 14% AMD PPA gain and Jalapeño area cuts

OpenAI reports up to 14% PPA gains on AMD GPU IP, 13% and 10% Jalapeño area reductions, and a 9,881-run RISC-V benchmark led by GPT-6.2 Astra at 112,000 tokens per run.

Advancing chip design with OpenAI models
Advancing chip design with OpenAI modelsAI-generated
By Rebecca Stone6 min read

OpenAI says its models helped reduce standard-cell area in its Jalapeño inference chip by about 13% in the vector unit and 10% in the matrix unit. The company also reports up to 14% PPA improvement over heavily optimized baselines on AMD’s flagship GPU IP. In OpenAI’s terminology, PPA covers power, performance and area. OpenAI presented the results as evidence that models can make improvements in production chip designs.

What changed inside Jalapeño?

OpenAI began using GPT‑o3 mini to help design Jalapeño in early 2025, according to the company. Jalapeño is OpenAI’s first custom inference chip. In an earlier post in August, OpenAI said engineers used frontier models to accelerate its development. The latest account focuses on expert-level design improvements and the same approach applied to third-party hardware with AMD.

OpenAI said even GPT‑o3 mini found opportunities for nontrivial PPA improvements. It attributed the early gains to changes in two arithmetic blocks:

  • In the vector unit, which manages and executes parallel computing tasks, OpenAI cited savings from floating-point adders and multipliers, conversion functions, reduction logic and the special-function unit.
  • OpenAI said the models explored alternative logic implementations and how to divide computations across pipeline stages. Those experiments found ways to share hardware and reduce intermediate storage.
  • In the matrix unit, an arithmetic component, OpenAI found improvements across floating-point adders, multipliers and low-precision dot-product circuits.
  • 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 retaining its two-cycle latency, according to OpenAI.
  • Further improvements to dot-product logic and pipeline organization produced additional savings across the matrix block, OpenAI said.

OpenAI estimated that model-proposed changes across Jalapeño reduced standard-cell area by about 13% in the vector unit and 10% in the matrix unit. The company said Jalapeño’s silicon bring-up proved the optimizations.

How did the models rewrite chip circuits?

OpenAI said the model reasoned “at the level of an expert microarchitect.” The company said the model rewrote circuits that perform calculations, control operations or move data.

OpenAI also said the model used Synopsys tools to verify that each rewrite preserved a block’s behavior and to evaluate trade-offs in power, performance and area. For AMD, OpenAI said the model identified further improvements by rewriting SystemVerilog RTL, the code that describes a chip’s circuits.

What did AMD verify?

According to OpenAI, AMD has already included some of these changes in a recent design it sent for manufacturing. AMD also added others to its production design files for future versions. That gives the reported gains production stakes beyond OpenAI’s own inference chip.

OpenAI also reported that AMD verified single-digit percentage gains in PPA across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. The company expects further PPA improvements on top of the GPU and CPU relative to the most optimized baseline.

OpenAI described the pace of the AMD work as “roughly a year of progress achieved in less than 5 months–unprecedented results at the frontier of chip design.”

How did OpenAI benchmark chip-design reasoning?

OpenAI evaluated successive model generations on two open-source 32-bit RISC-V processor cores. PicoRV32 is a compact core designed for reprogrammable FPGAs. OpenTitan uses Ibex as the embedded processor in its security chip. OpenAI described both as complete processor cores for embedded, microcontroller-class workloads.

OpenAI gave each model the original hardware description for each core and asked it to increase throughput, the number of instructions a core can execute per second. The company said the measure captured both an instruction finishing in fewer clock cycles and changes that made each cycle take longer.

OpenAI’s evaluation used a fixed set of checks and constraints:

  • OpenAI assigned a zero to designs that failed a check or used more than 16 times the original core’s area.
  • For designs that passed, OpenAI measured instructions per clock cycle on workloads hidden from the models.
  • The company took the geometric mean across those workloads and multiplied it by the maximum clock frequency estimated with Yosys, producing an estimate of instructions per second.
  • For each core, OpenAI assigned the best throughput observed under the same constraints a score of 1.0.
  • The charts showed average scores across runs, including runs that scored zero.

With a comparable budget of about 112,000 tokens per run, GPT‑6.2 Astra had the highest average score on both cores. GPT‑6.1 Sol also outscored GPT‑5.6 Sol on both.

What did the from-scratch task reveal?

OpenAI then gave each model written requirements and templates from the open-source code showing how each core connects to the rest of the system. The model chose how data would move through the core and how to divide the work into stages. OpenAI kept the correctness checks, memory and cache settings, and throughput objectives the same as in the RTL optimization task.

At about 112,000 tokens per run, all five models scored higher when they designed a core from written requirements than when they tried to improve the existing design. That result held for both PicoRV32 and Ibex. GPT‑6.2 Astra gained more on Ibex than on PicoRV32.

The full evaluation included 9,881 scored runs across both tasks, both cores and all token budgets.

What caveats limit the RISC-V results?

OpenAI said starting from written requirements gives models room to omit features found in the original cores, including performance counters and additional instructions. Those features may do little to improve throughput on the test workloads, OpenAI said, which could explain part of the score advantage.

Requiring those features, setting a tighter limit on circuit area and checking timing after physical synthesis would make the test more representative of production chip design, according to OpenAI. The company says the RISC-V results show that models can explore new processor architectures. It separately says the Jalapeño and AMD work added net-new improvements to already sophisticated production designs.

What changes next with Synopsys?

OpenAI and Synopsys are working to integrate frontier intelligence with electronic design automation tools that engineers already use to design and verify chips. OpenAI says the partnership will train GPT‑Synopsys, a new model, to use Synopsys EDA tools the way expert chip designers do. The company said the collaboration also aims to ground model development in trusted evaluators and broaden the design challenges addressed.

OpenAI says the current RISC-V test still lacks three production checks: required original features, tighter area limits and timing after physical synthesis. Those checks define the next test for GPT‑Synopsys: whether it can extend the reported gains to more representative production designs.

Source: OpenAI News

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

754 articles

Related articles

  1. OpenAI says its models cut AMD flagship GPU IP power-performance gaps by 14%
  2. OpenAI Posts First Benchmark Results for Jalapeño Custom Inference Chip
  3. OpenAI's Jalapeño Chip: LLMs Cut Design Time to Record Lows
  4. OpenAI Reports Up to 14% PPA Gains in AMD Chip Designs
  5. OpenAI's Jalapeño Chip Beats Rivals on Power and Latency

« Previous article