
OpenAI Claims Up to 14% PPA Gains on AMD's Flagship GPU IP
OpenAI says its models delivered up to 14% PPA gains on AMD's flagship GPU IP in under five months, with some changes already sent for manufacturing. AMD has not commented.
- By
- Nathan Brooks
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
OpenAI reported gains of up to 14 percent on power, performance and area across AMD's flagship GPU IP, a combined chip-design measure known as PPA, achieved in under five months of work. The company said on Nov. 18 that its models reached the figure by rewriting the SystemVerilog code that describes the chip's circuits. OpenAI frames this as roughly a year of engineering progress compressed into less than half of one.
The claim carries caveats worth reading closely. The phrase "up to 14% improvement relative to heavily optimized baselines" describes a maximum, and OpenAI does not break the number down by power, speed or area. The announcement post also contains no statement from AMD. The gains are OpenAI's account of OpenAI's model, benchmarked against baselines OpenAI characterizes as heavily optimized. That is a claim, not a tape-out.
What has actually shipped?
Some of the story is concrete. According to OpenAI, AMD has already folded some model-proposed changes into a recent design it sent for manufacturing. Others have landed in AMD's production design files for future versions. That is the difference between a paper result and silicon on a wafer, and AMD has crossed at least part of that line.
The CPU side is thinner but real. OpenAI said AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU, using designs written by the model. No block-level detail accompanied the figure.
Jalapeño and the Synopsys project
OpenAI's first custom inference chip, Jalapeño, supplies the internal test case. Model-proposed changes cut standard-cell area by about 13 percent in the chip's vector unit and 10 percent in its matrix unit. Area is the easiest PPA axis to verify and the hardest to game with measurement choices, which makes these numbers more digestible than the headline figure.
On two open-source RISC-V processor cores, GPT-6.2 Astra posted the highest average score at a comparable budget of roughly 112,000 tokens per run. OpenAI is also training GPT-Synopsys with Synopsys itself, a model that works with the EDA vendor's design tools. If a language model learns to drive the tools that build chips, the interesting question becomes who owns the resulting productivity gain and at what license price.
Why the commercial ties matter
None of this happens between neutral parties. Under a deal announced in October 2025, OpenAI agreed to deploy six gigawatts of AMD GPUs and received the right to buy up to 160 million AMD shares at a low fixed price. That could amount to a stake of roughly 10 percent, with vesting tied to deployment milestones and AMD share-price targets. AMD, in July, described the pair as "jointly researching how AI can improve the systems on which AI runs."
So the lab result doubles as an equity story. A design partnership in which one party holds warrants on the other's stock has a built-in incentive to publish flattering numbers. It does not make the numbers wrong; it makes them unilaterally sourced.
The unanswered question
The "up to 14%" figure will circulate stripped of its qualifiers, as such figures do. What would settle the matter is an AMD-authored statement, a named test case and a breakdown across power, performance and area. Until then, the verified facts are modest: some changes are in manufacturing, some are in design files for future parts, and single-digit gains hold on a next-generation CPU. Whether model-written RTL becomes standard practice or remains a benchmark curiosity depends on the next tape-out, not the next blog post.
Source: mlq.ai
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,092 articles
Related articles
openai-reportedly-plans-to-buy-google-tpus-for-ai-compute
OpenAI reportedly plans to buy Google TPUs for AI compute
openai-turns-to-samsung-for-next-gen-ai-chips-as-partnership-deepens-76564f5f
OpenAI Turns to Samsung for Next-Gen AI Chips as Partnership Deepens
altman-flags-ai-chip-demand-concentration-as-capacity-expands
Altman Flags AI Chip Demand Concentration as Capacity Expands
openai-weighs-google-tpus-as-it-works-to-cut-its-nvidia-dependence
OpenAI Weighs Google TPUs as It Works to Cut Its Nvidia Dependence



