OpenAI reported up to 14% improvement in power, performance, and area (PPA) for AMD flagship GPU intellectual property, compared with heavily optimized baselines. The figure describes a combined design metric; OpenAI also said some model-proposed changes were included in a design sent for manufacturing. OpenAI’s announcement was published November 18, 2026.

What OpenAI’s 14% figure covers

PPA combines power, performance, and area to assess design trade-offs. OpenAI’s reported maximum applies to AMD flagship GPU IP and its comparison with heavily optimized baselines. It is a result about chip design, rather than a measured performance figure for a finished GPU.

OpenAI did not identify the GPU design or specify which changes went into the manufacturing-bound design. The company said other proposed changes were added to production design files for future versions.

The key detail is the scope: OpenAI’s percentage applies to a combined PPA measure for AMD GPU IP.

How the model-assisted changes were checked

OpenAI says its models rewrote SystemVerilog register-transfer-level (RTL) descriptions—the code that specifies a circuit’s operations and data movement. Synopsys tools checked whether the rewrites preserved block behavior and evaluated their power, performance, and area trade-offs, according to OpenAI.

That workflow connects code changes to engineering checks before design changes move forward. OpenAI said the work took less than five months, building on years of AMD engineering.

Lisa Su’s public reaction

On November 18, AMD CEO Lisa Su praised the engineering collaboration in a public post. Her post was a statement of support; OpenAI’s announcement supplied the reported 14% figure.