Particle.news

OpenAI Says Its AI Improved AMD GPU Designs by Up to 14%

The company says model-generated RTL changes have been validated in silicon and moved into manufacturing while it builds a specialised EDA model with Synopsys to scale the workflow.

Overview

  • OpenAI disclosed Thursday that its generative models produced up to 14 percent improvements in performance, power, and area on AMD’s flagship GPU intellectual property and that some of those AI-generated changes were included in designs sent for manufacturing.
  • The company reported single-digit gains on critical blocks of AMD’s next-generation CPU and said model edits reduced standard-cell area by about 13 percent in Jalapeño’s vector unit and 10 percent in its matrix unit, with those Jalapeño changes validated during silicon bring-up.
  • OpenAI used models to rewrite SystemVerilog RTL and ran Synopsys electronic design automation tools to check behavioral equivalence and to measure power, performance, and area before moving changes into engineers’ flows.
  • OpenAI announced a partnership with Synopsys to train a specialised GPT-Synopsys model to operate EDA tools like an expert designer, and Synopsys said it will embed OpenAI intelligence directly into EDA workflows.
  • OpenAI warned that designs generated from written specs can omit expected elements such as counters and need tighter area and timing checks and standard verification before AI-proposed changes can be treated as production-ready, signaling engineers must keep conventional safeguards.