New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.
Get Free API Key
Semiconductors & AI Research•4 min read•2026-11-20

OpenAI Models Rewrite Silicon Circuits: 14% PPA Gain on AMD Flagship GPU IP and GPT-Synopsys Partnership

OpenAI reveals its AI models achieved up to 14% PPA gains on AMD flagship GPU IP, shrank Jalapeno chip area by up to 13%, and partnered with Synopsys to train GPT-Synopsys.

Tech News Bot
DeviceSpecs Intelligence
Original Source Reference
#OpenAI#AMD#Synopsys#Chip Design#Jalapeno#GPT-6.2 Astra#RISC-V#Semiconductors

OpenAI has published new research demonstrating that its frontier AI models are generating net-new improvements in production semiconductor engineering by rewriting SystemVerilog RTL circuits. In collaboration with AMD, OpenAI models delivered up to 14 percent power, performance, and area (PPA) gains on flagship GPU IP in under five months, while shrinking compute unit area on OpenAI in-house Jalapeno inference chip and launching a co-training partnership with Synopsys.

14% PPA Gains on AMD Flagship GPU IP and CPU Blocks

Working directly with AMD engineering teams, OpenAI models analyzed and rewrote SystemVerilog RTL descriptions across heavily optimized production blocks. The models achieved up to 14 percent PPA improvement on AMD flagship GPU IP in less than five months—compressing roughly a year of traditional microarchitectural tuning. AMD also verified single-digit percentage PPA gains on key blocks of an upcoming next-generation flagship CPU, folding several AI-authored optimizations directly into designs already sent to manufacturing.

Silicon-Validated Area Reductions on OpenAI Jalapeno Chip

OpenAI detailed how it began deploying models like GPT-o3 mini in early 2025 to optimize its custom Jalapeño AI inference silicon. By refining floating-point adders, multipliers, conversion functions, and low-precision dot-product circuits, model-proposed RTL changes reduced standard-cell area by approximately 13 percent in the vector unit and 10 percent in the matrix unit—efficiency improvements that have since been validated in physical silicon.

RISC-V Benchmarks and the GPT-Synopsys EDA Partnership

To measure generational progress, OpenAI ran 9,881 scored evaluations optimizing and synthesizing two open-source 32-bit RISC-V cores, PicoRV32 and OpenTitan Ibex. Around a 112,000-token budget, GPT-6.2 Astra ranked first on average, scoring roughly double GPT-6 Astra on PicoRV32 optimization. Moving toward full production toolchains, OpenAI also announced a partnership with electronic design automation leader Synopsys to co-train GPT-Synopsys, equipping AI agents to operate professional EDA suites.