Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

OpenAI's Models Rewrote Parts of AMD's Flagship GPU Design for Up to 14% Better Power, Performance and Area — and Some Changes Are Already Taped Out

OpenAI's Models Rewrote Parts of AMD's Flagship GPU Design for Up to 14% Better Power, Performance and Area — and Some Changes Are Already Taped Out

OpenAI says its models rewrote circuit code on AMD's flagship GPU for up to 14 percent better power, performance and area in under five months, and AMD has already sent some changes for manufacturing. CEO Lisa Su called the results unprecedented. On OpenAI's Jalapeño inference chip, model-proposed changes cut the warp engine by about 13 percent and matrix unit by about 10 percent, both proven on silicon. All figures are company-reported.

OpenAI said on Tuesday that its models redesigned portions of AMD's flagship GPU intellectual property, delivering up to 14 percent better power, performance and area against heavily optimized baselines in under five months — and that AMD has already moved some of the AI-proposed changes into a design it recently sent for manufacturing. The results were published November 18 in a technical post titled "Advancing chip design with OpenAI models."

The 14 percent figure applies to individual blocks of the GPU design rather than the finished processor, and OpenAI describes it as roughly a year of engineering progress compressed into less than five months. AMD CEO Lisa Su publicly congratulated the joint engineering teams, calling the outcome "fantastic" and "unprecedented results at the frontier of chip design," and confirmed that part of the work is already inside a recently taped-out design. All performance claims in the announcement are company-reported and have not been independently verified.

The models worked directly on SystemVerilog RTL — the hardware description code that defines how circuits compute, control and move data. According to OpenAI, the models reasoned at the level of an expert microarchitect, and every proposed rewrite was validated in Synopsys EDA tools to confirm the block's behavior was unchanged before AMD engineers reviewed it for power, performance and area trade-offs. AMD separately verified single-digit percentage PPA gains across key blocks of its upcoming flagship CPU using model-written designs.

The same workflow produced silicon-proven gains on Jalapeño, OpenAI's first custom inference chip, a design the company began optimizing with GPT-4o in early 2025. OpenAI estimates model-proposed changes cut standard-cell area by about 13 percent in the warp engine and about 10 percent in the matrix unit. A redesigned multiplier uses roughly 15 percent less cell area and now completes in a single clock cycle instead of two, and an adder redesign saved about 15 percent of its area while keeping two-cycle latency. In an August post on Jalapeño's first results, OpenAI reported 1.5 to 1.9 times more AI work per watt at peak throughput than comparison systems, with deployment in its own infrastructure planned by the end of the year.

OpenAI also published a capability evaluation built on two open-source RISC-V cores, PicoRV32 and OpenTitan's Ibex, covering 9,881 scored runs across optimization and design-from-specification tasks. At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra posted the highest average score — roughly twice GPT-6 Astra's on PicoRV32 — and GPT-6.1 Sol outscored GPT-5.6 Sol. OpenAI notes models scored higher when designing from a written specification, partly because they may omit rarely used features such as performance counters.

The announcement extends the GPT-Synopsys partnership OpenAI and Synopsys disclosed on September 30 — a specialized model being trained to operate Synopsys' EDA design and validation tools the way expert chip designers do, with early customer engagements under way.

The work deepens a commercial relationship that already runs through silicon supply: in October 2025, AMD and OpenAI signed a multi-year agreement to deploy 6 gigawatts of AMD Instinct GPUs, with the first 1-gigawatt tranche of MI450 hardware due to begin deployment in the second half of 2026.

If the reported gains hold up in production, the significance goes beyond a single vendor's benchmark. Frontier labs are starting to point their own models at the most expensive, slowest-moving part of the AI stack — chip development — where a year of schedule compression compounds across every future GPU generation. For AMD, an AI-tuned design flow offers a way to compete on efficiency rather than raw transistor count; for Synopsys and the wider EDA industry, AI is becoming the design flow itself rather than merely a customer of it.

Comments (0)

Log in to join the discussion

Log In

No comments yet