> ## Content Index
> Fetch the complete content index at: https://www.tao.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI Reports Up to 14% Gains in AMD GPU Design and Smaller Jalapeño Circuits
- URL: https://www.tao.media/openai-reports-up-to-14-gains-in-amd-gpu-design-and-smaller-jalapeno-circuits/
- Published: 2026-11-19T17:21:36.000Z
- Updated: 2026-11-19T17:21:37.000Z
- Description: Model-written circuit changes are entering AMD designs, while OpenAI reports silicon-tested area savings in its own inference chip.
- Author: Bart Hillerich
- Tags: OpenAI, AMD, Synopsys, AI, News

OpenAI says its models found improvements of up to 14% in power, performance, and area on AMD's flagship GPU intellectual property, with some model-written changes already incorporated into a design sent for manufacturing. The work took less than five months and built on circuits AMD engineers had spent years optimizing.

The collaboration extends an approach OpenAI used on Jalapeño, its first custom inference chip. OpenAI estimates that model-proposed changes reduced standard-cell area by about 13% in Jalapeño's vector unit and 10% in its matrix unit, with the optimizations verified during testing of the manufactured silicon.

The results detailed in OpenAI's [chip-design research update](https://openai.com/index/advancing-chip-design-with-our-models/?ref=tao.media) describe model-written circuit changes entering production hardware. The workflow combines AI-written hardware descriptions with established engineering tools that check whether a rewrite preserves the circuit's behavior and improves its implementation.

![OpenAI and AMD chip design results, image 1 of 4](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HXbDvvBbkAA_huj.jpg)

[OpenAI and AMD chip design results, image 1 of 4 | OpenAI (X)](https://x.com/OpenAI/status/2123100386638877054?ref=tao.media)

## Rewriting AMD Circuits With Verification Tools

The models worked in SystemVerilog RTL, or register-transfer level code, which describes how a chip performs calculations, controls operations, and moves data. Unlike software that runs on a processor, RTL defines the circuitry from which the processor itself is built.

Changing that code can alter how much logic a circuit requires, how quickly it operates, and how much power it consumes. Those competing objectives make up PPA: power, performance, and area. A smaller circuit, for example, is useful only if it still meets the design's functional and timing requirements.

In the AMD collaboration, Synopsys tools checked that each rewrite preserved the block's behavior and evaluated the resulting PPA trade-offs. OpenAI says the work delivered roughly a year's worth of additional design progress in less than five months. Its reported maximum of 14% applies to flagship GPU IP relative to heavily optimized baselines, rather than a measured 14% speedup for an entire GPU.

![OpenAI and AMD chip design results, image 2 of 4](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HXbDvvybIAAlN65.jpg)

[OpenAI and AMD chip design results, image 2 of 4 | OpenAI (X)](https://x.com/OpenAI/status/2123100386638877054?ref=tao.media)

According to OpenAI, AMD has put some changes into a recent design sent for manufacturing and added others to production design files for future versions. OpenAI also reports that AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using model-written designs.

The work therefore spans both graphics and general-purpose processors, while retaining conventional verification and evaluation tools as the checks on proposed circuit changes.

## Where Jalapeño Saved Circuit Area

OpenAI began using GPT-o3 mini to help design Jalapeño in early 2025\. The company's [first Jalapeño results](https://openai.com/index/jalapeno-first-results/?ref=tao.media), published in August, described how AI assisted chip development and software optimization. The new results explain where the hardware-level savings came from.

In the vector unit, the models improved floating-point adders and multipliers, conversion functions, reduction logic, and the special-function unit. These circuits handle arithmetic and related operations within the chip's parallel computing hardware.

The models explored alternative logic implementations and different ways to divide calculations across pipeline stages. Pipelining splits work into steps separated by storage, allowing successive operations to move through the circuit. Changing those boundaries can create opportunities to share hardware or reduce the intermediate values that must be stored.

The matrix unit's improvements accumulated across floating-point arithmetic and low-precision dot-product circuits, which combine multiplications and additions used in AI computation. OpenAI reports that one multiplier redesign used about 15% less cell area in synthesis while reducing latency from two clock cycles to one. A separate adder redesign reduced cell area by about 15% while keeping its two-cycle latency.

![OpenAI and AMD chip design results, image 3 of 4](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HXbDvvnakAAzZqh.jpg)

[OpenAI and AMD chip design results, image 3 of 4 | OpenAI (X)](https://x.com/OpenAI/status/2123100386638877054?ref=tao.media)

Across the broader effort, OpenAI's team estimates savings of about 13% in vector-unit standard-cell area and 10% in matrix-unit standard-cell area. Standard cells are the building blocks used to implement digital logic; these figures describe the affected units, not a corresponding reduction in total chip size. OpenAI says the optimizations were silicon-proven during Jalapeño's bring-up, when engineers tested the manufactured chip.

## Testing New Processor Designs

OpenAI also evaluated successive model generations on two open-source, 32-bit RISC-V processor cores: PicoRV32, a compact design for programmable FPGA hardware, and Ibex, the embedded processor used in OpenTitan's security chip.

One task asked models to improve existing RTL. Another provided written requirements and interface templates, leaving models to choose how data moved through the processor and how work was divided into stages.

OpenAI measured throughput by combining instructions per clock cycle on hidden workloads with a maximum clock frequency estimated using Yosys. Designs received zero if they failed a simulation check or exceeded 16 times the original core's area. Scores were normalized separately for each core, with the best observed throughput under the same constraints assigned 1.0.

At a comparable budget of about 112,000 tokens per run, OpenAI reports that [GPT-6.2 Astra](https://tao.media/openai-launches-gpt-6-2-astra-and-sol-cutting-the-cost-of-professional-ai-work/?ref=tao.media) achieved the highest average score on both cores. GPT-6.1 Sol also outperformed GPT-5.6 Sol on both. The full evaluation covered 9,881 scored runs across the two tasks, both cores, and all token budgets.

![OpenAI and AMD chip design results, image 4 of 4](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/HXbDvufasAA5M5N.jpg)

[OpenAI and AMD chip design results, image 4 of 4 | OpenAI (X)](https://x.com/OpenAI/status/2123100386638877054?ref=tao.media)

All five tested models scored higher when designing from written requirements than when modifying existing cores at roughly the same token budget. OpenAI notes that starting from requirements allowed models to omit features such as performance counters and additional instructions, potentially explaining part of that advantage. Tighter area limits, required feature parity, and timing checks after physical synthesis would make the evaluation more representative of production design.

## Training a Model to Use Chip-Design Tools

OpenAI and Synopsys are developing [GPT-Synopsys](https://investor.synopsys.com/news/news-details/2026/OpenAI-and-Synopsys-Announce-GPT-Synopsys-Frontier-Intelligence-to-Revolutionize-Chip-Design/default.aspx?ref=tao.media), a specialized model intended to operate Synopsys electronic design automation tools, interpret their outputs, and iterate toward verified designs for engineer review.

The partnership was announced in September under a multiyear agreement covering research, tool licensing, and joint commercialization. The planned service will bundle compute, the model, and tool licenses, run on OpenAI-hosted infrastructure, and integrate with Synopsys.ai and Synopsys Autopilot. Synopsys says early technology engagements with semiconductor customers are underway.