AI-Written Circuit Changes Entered an AMD Design Sent for Manufacturing
OpenAI says AMD accepted model-proposed RTL changes into a design sent for manufacturing after reporting up to 14% block-level PPA improvement. Separate Jalapeño optimizations were validated in silicon, but the headline figure is company-reported and does not describe whole-chip performance.
Quick summary
- OpenAI says model-proposed RTL edits improved the combined power, performance and area score of selected AMD GPU blocks by as much as 14% against already optimized baselines.
- AMD incorporated some of those changes into a design sent for manufacturing, while separate Jalapeño processor optimizations have already been validated in silicon.
- The figures are block-level company results, not proof that an entire GPU is 14% faster.
Artificial intelligence has crossed a consequential boundary in chip design: from suggesting experiments to contributing circuit changes that reached manufacturing files. OpenAI reported on November 18 that AMD engineers used its models to optimize register-transfer level, or RTL, descriptions inside flagship processor programs.
The headline result was an improvement of up to 14% in a composite measure of power, performance and area, known as PPA, on selected blocks of AMD flagship GPU intellectual property. OpenAI says the comparison used heavily optimized production baselines rather than classroom designs. Some model-proposed edits were accepted into an AMD design that was sent for manufacturing, and other edits entered production design files.
That is more meaningful than a benchmark win, but narrower than a whole-chip performance claim. It shows that human-reviewed AI proposals can survive industrial verification and enter a real hardware workflow.
What the models changed
RTL is the layer where engineers describe how data moves between registers and how logic operates each clock cycle. A change at this level can alter timing, cell count and energy use after synthesis. OpenAI says AMD also confirmed single-digit PPA gains in important blocks of a next-generation flagship CPU.
The clearest silicon-backed example came from Jalapeño, OpenAI's inference processor. Model-assisted edits reduced standard-cell area by about 13% in a vector unit and 10% in a matrix unit. One multiplier used roughly 15% less cell area while cutting latency from two clock cycles to one, and an adder was also reduced by about 15%. These blocks were subsequently exercised during silicon bring-up.
The process did not let a language model approve its own work. Synopsys tools performed deterministic checks of functional behavior and implementation quality, while engineers reviewed whether the proposals preserved the intended specification. This pairing matters because a plausible-looking RTL rewrite can still omit a feature, alter an edge case or merely shift costs elsewhere.
Evidence beyond one proprietary design
OpenAI also disclosed a reproducible evaluation using the open-source PicoRV32 and Ibex RISC-V cores. Across 9,881 runs, its strongest model produced the best comparable mean result in the test set. Yet the company explicitly notes limitations: specifications can be incomplete, generated designs were allowed to use up to 16 times the baseline area, and clock frequency was estimated before full physical synthesis.
The commercial context is expanding. OpenAI and Synopsys announced a multiyear partnership to place OpenAI models inside engineering workflows spanning chips, systems and software. That relationship helps explain the verification stack, but it also means the principal evidence comes from parties selling or adopting the technology.
Related NewTqnia coverage shows the wider transition from research circuits to production toolchains. A hybrid MoS2 and carbon-nanotube CMOS adder illustrates how device-level advances are measured, while AWS opening Trainium to native PyTorch shows how chip adoption also depends on software accessibility.
Reality check
The 14% figure is a company-reported maximum for selected blocks and a composite PPA metric. OpenAI and AMD have not disclosed the exact baseline, affected AMD product, foundry process or whole-chip impact. A design being sent for manufacturing is not the same as successful silicon. Independent confirmation will require shipped hardware, repeatable measurements and disclosure of which AI-authored changes survived final implementation.
Verified topics and entities
Sources and citations4 sources
External references used to support the reporting in this article.
Published by
NewTqnia Artificial Intelligence Desk
An institutional editorial team within NewTqnia