Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
InfrastructureInsight · On RadarThe Decoder→ original

OpenAI's AI Improved an AMD GPU Design by Up to 14% in Power, Speed and Area in 5 Months

OpenAI says its models helped AMD improve a flagship GPU block by up to 14% on PPA (power, performance and die area) in under five months. By OpenAI's estimate, that equals about a year of work on top of a design AMD engineers had polished for years. Some of the changes have already made it into a design sent to manufacturing. For AI products, this could mean cheaper inference, though the figures so far come only from the company itself.

AI-processed from The Decoder; edited by Hamidun News
OpenAI's AI Improved an AMD GPU Design by Up to 14% in Power, Speed and Area in 5 Months
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1The same method helped with Jalapeño, OpenAI's first in-house inference chip
Key metric #2On open RISC-V cores, GPT-6.2 Astra delivered the best result
Key metric #3Anthropic's Mythos processed prompts on AMD MI355X in 3 days instead of 6–12 months

Imagine a codebase that a team has optimized for years: profilers, refactoring, manual tuning. Then a neural network finds up to 14% more headroom in it. That is roughly what OpenAI now claims about chips. Together with AMD, its models improved a flagship GPU block by up to 14% on PPA in under five months. PPA covers the three parameters chip engineers are always fighting over: power consumption, performance and die area. The better the PPA, the faster, more efficient or more compact the chip. This matters to anyone who pays to run AI models, because hardware determines the cost of every response. The data comes from OpenAI itself, and there is no independent verification yet.

What exactly OpenAI and AMD achieved

The work concerns a so-called IP block: a ready-made piece of chip design that is later used in products. This is a flagship AMD GPU block, not a rough draft where any new approach quickly finds slack. AMD engineers have been optimizing it for years.

By OpenAI's estimate, an improvement of up to 14% in PPA in less than five months equals about a year of progress on top of the work humans had already done. Two caveats matter here: "up to 14%" is a maximum, not an average, and "a year of progress" is OpenAI's own estimate, not a measured value.

There is also a practical detail: according to OpenAI, AMD has already folded some of the changes into a design sent to manufacturing. That separates this from a lab demo: the changes go through a process where a mistake is very expensive.

Jalapeño, RISC-V and Synopsys

The same method helped with Jalapeño, OpenAI's first in-house chip for inference, meaning running already-trained models. For a company that buys compute in enormous volumes, every watt and every square millimeter of silicon saved translates directly into money.

Separately, OpenAI tested its models on open RISC-V processor cores. There, the new GPT-6.2 Astra showed the best results of any model generation. For practitioners, this is a signal that models' ability to help with chip design improves from version to version.

OpenAI is also working with Synopsys on GPT-Synopsys, a model that uses Synopsys's electronic design automation (EDA) tools. The goal is to make AI-assisted chip design accessible to a wider range of teams.

Anthropic is taking the same path

Anthropic reports similar results. According to the company, its Mythos model processed the prompts for a model on AMD MI355X GPUs in just three days, while such a task normally takes from six to twelve months. Claude currently runs on four chip families, which the company uses to optimize cost and performance per dollar.

What this means

For founders and product managers, the takeaway is pragmatic. Inference is the largest cost for many AI products, and a gain of a few percent at the chip level can eventually turn into lower prices or higher margins. If adapting models to new hardware shrinks from months to days, as in Anthropic's example, lock-in to a single GPU vendor becomes less rigid.

Caution is also warranted. It is unknown how much human oversight was needed to verify the proposed changes, or how well the results will carry over to other blocks. But the direction is clear: AI is starting to speed up the creation of the hardware that the next generation of AI runs on.

Story

Key Questions Answered

What exactly OpenAI and AMD achieved?

The work concerns a so-called IP block: a ready-made piece of chip design that is later used in products. This is a flagship AMD GPU block, not a rough draft where any new approach quickly finds slack. AMD engineers have been optimizing it for years. By OpenAI's estimate, an improvement of up to 14% in PPA in less than five months equals about a year of progress on top of the work humans had already done.

What are the key details on Jalapeño, RISC-V and Synopsys?

The same method helped with Jalapeño, OpenAI's first in-house chip for inference, meaning running already-trained models. For a company that buys compute in enormous volumes, every watt and every square millimeter of silicon saved translates directly into money. Separately, OpenAI tested its models on open RISC-V processor cores. There, the new GPT-6.2 Astra showed the best results of any model generation.

What are the key details on Anthropic is taking the same path?

Anthropic reports similar results. According to the company, its Mythos model processed the prompts for a model on AMD MI355X GPUs in just three days, while such a task normally takes from six to twelve months. Claude currently runs on four chip families, which the company uses to optimize cost and performance per dollar.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…