AIO Apex
en

OpenAI's AI Models Rewrote AMD GPU Circuits, Delivering Up to 14% Efficiency Gains

OpenAI
Share:
OpenAI's AI Models Rewrote AMD GPU Circuits, Delivering Up to 14% Efficiency Gains

OpenAI announced on November 18 that its AI models have redesigned AMD's flagship GPU circuit IP, achieving up to 14% improvement in power, performance, and area (PPA) metrics — work that AMD engineers estimate would have taken roughly a year. Some of those redesigned circuits are already inside a chip AMD has submitted to a fabrication facility.

The models worked at the RTL (register-transfer level) layer, rewriting SystemVerilog code that describes the logical structure of GPU circuits. AMD verified the results using Synopsys electronic design automation tools, confirming that model-proposed changes preserved the original chip functionality while delivering measurable PPA gains. Beyond GPUs, AMD also verified single-digit PPA improvements across key blocks of an upcoming next-generation flagship CPU.

OpenAI simultaneously reported gains on its own Jalapeño inference chip, which the company uses internally to serve its models. Greg Brockman, President of OpenAI, described the AMD work as «roughly an additional year of chip design progress accomplished in under five months.»

The 14% figure refers to improvements in specific GPU circuit blocks, not a whole-chip speedup. PPA — the standard semiconductor design metric — measures how well a circuit balances energy efficiency (power), clock speed (performance), and silicon real estate (area). Gains at the block level compound across a chip but do not translate directly to finished product benchmarks. Some model-proposed changes are already incorporated into a GPU design AMD submitted for manufacturing; others are staged in production design files for future chip generations.

As part of the same initiative, OpenAI and Synopsys are developing a new EDA model — called GPT-Synopsys — trained to operate within the Synopsys professional chip design toolchain. This marks a step toward AI models that understand the full semiconductor design workflow rather than just generating RTL code.

Human engineers review every model-proposed change before it ships. OpenAI's chip design team reportedly does not yet fully trust AI-generated circuit modifications, keeping engineers as the final verification layer. That conservatism likely explains why the results are credible: no model change has bypassed human sign-off.

The work builds on the multi-year AMD-OpenAI compute partnership announced in October 2025, under which AMD committed 6 gigawatts of GPU capacity across its Instinct GPU generations. The chip design collaboration was not originally disclosed as part of that agreement and represents a deeper technical integration than the supply deal alone implied.

As first reported by OpenAI in its blog post «Advancing Chip Design with OpenAI Models,» published November 18, 2026.

Originally reported by OpenAI. Read the original article for additional details.

View original source
Share: