Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

OpenAI says its models delivered up to 14% gains on AMD's flagship GPU in under five months

OpenAI says its models delivered up to 14% gains in power, performance and area on AMD's flagship GPU IP in under five months, work it describes as roughly a year of chip-design progress. AMD has already folded some changes into a design sent for manufacturing, while OpenAI's new GPT-6.2 Astra led its RISC-V design evaluations.

Published

On November 18, OpenAI published an engineering post describing work with AMD in which its models delivered up to 14% improvement in power, performance and area (PPA) on AMD's flagship GPU IP, measured against already heavily optimized baselines. OpenAI says the result amounts to roughly a year of chip-design progress achieved in less than five months.

The method is not a chatbot writing advice. According to OpenAI, the model rewrote SystemVerilog RTL, the code that describes a chip's circuits, touching logic that performs calculations, controls operations and moves data between blocks. OpenAI says it reasoned about RTL at the level of an expert microarchitect, and that each rewrite was checked with Synopsys tools to confirm the block's behavior was preserved and to weigh the trade-offs in power, performance and area.

Some of that work has already left the lab. AMD has included part of the changes in a recent design it sent for manufacturing and added others to production design files for future versions, OpenAI says. AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using designs written by the model.

The same approach was applied to Jalapeño, OpenAI's first custom inference chip, where the company says it began working with GPT-4o in early 2025. Model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit; one multiplier redesign used about 15% less cell area while dropping latency from two clock cycles to one, and an adder redesign saved about 15% area at the same two-cycle latency. Those optimizations were silicon-proven during Jalapeño's bring-up.

To track how capability changes across model generations, OpenAI ran two open-source 32-bit RISC-V cores, PicoRV32 and the Ibex core used in OpenTitan, through an optimization task: models were given the original hardware description and asked to raise throughput, with correctness and performance checked on hidden workloads. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score on both cores, and GPT-6.1 Sol beat GPT-5.6 Sol on both. The full evaluation covered 9,881 scored runs.

The significance is that models are moving from coding assistants to participants in production microarchitecture. Chip design is bounded by scarce senior engineer time, with each block's timing, area and power settled through repeated trade-offs, and a model that can generate many verifiable rewrites in hours shifts human effort toward judgment and verification.

OpenAI is also candid about the limits of its evaluation. Designing a core from a written specification lets a model omit features such as performance counters and extra instructions that do little for throughput on the test workloads, which may partly explain the higher scores; a stricter test would require those features, cap area more tightly and check timing after physical synthesis.

Two things are worth watching next. First, how much of the improvement now sitting in AMD's production design files survives into shipping GPUs and CPUs. Second, whether GPT-Synopsys, the model OpenAI is training with Synopsys to use EDA tools the way expert designers do, turns this kind of work from an internal experiment into a broadly usable engineering workflow.

Why it matters

If the improvements hold in shipping silicon, AI becomes more than a design assistant: it compresses design cycles and feeds directly into hardware competitiveness. For AMD that is an extra year of optimization on an existing architecture; for OpenAI it is a push into the core of semiconductor engineering.

OpenAIAMDChip Design
Back to realtime news

Nearby Updates

All

11/19, 02:52

Microsoft ships seven in-house MAI models at Ignite to replace outside models with cheaper alternatives

Microsoft unveiled seven in-house MAI models at Ignite spanning code, security, images, speech and routing, and is moving its own products onto them at a fraction of outside-model prices. Microsoft says MAI Code 2 reaches 98 percent of GPT-6.1 Sol's performance at 40 percent of the cost, framing the launch around token economics rather than benchmark leadership.

11/19, 02:04

Google gives AI Mode in Search its own incognito-style temporary chats

Google has added temporary threads to AI Mode in Search, letting people hold conversations that are not saved to search history or used to shape future recommendations. The feature brings an incognito-style option to generative search, addressing concerns about what conversational queries leave behind.

11/19, 02:03

Google opens its Spark AI agent to free users under the name “Tasks”

Google has opened its Spark AI agent to free users, presenting the capability under the name “Tasks,” according to a report from The Tech Buzz. The shift extends agentic, multi-step assistance beyond paying tiers to a much broader base of ordinary accounts.

11/19, 02:01

Sustainalytics flags high ESG risk at Anthropic, putting AI governance in focus

A Sustainalytics assessment carried by Morningstar places Anthropic in a high ESG risk category, tying the rating to questions about AI governance. The move shows how conventional investment analysis is starting to evaluate model developers on how they govern their own technology.