OpenAI Models Are Now Redesigning AMD Chips. Synopsys Is the Referee
AI improving the chips that run AI used to be a slide in a keynote. On Nov 18 OpenAI put numbers on it, and one of them comes from a customer: https://openai.com/index/advancing-chip-design-with-our-models
Start with OpenAI's own inference chip, JalapeΓ±o. The team began using GPT-4o on the design in early 2025, and by OpenAI's estimate model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, silicon-proven at bring-up. One multiplier redesign used about 15% less area and dropped latency from two clock cycles to one. Then AMD: OpenAI reports up to 14% PPA improvement over heavily optimized baselines on AMD's flagship GPU IP, which it calls roughly a year of progress in under five months. AMD has put some of the changes into a design already sent for manufacturing, and AMD itself verified single-digit PPA gains on key blocks of its next flagship CPU.
The part worth copying is not the model but the referee. Every SystemVerilog rewrite was checked with Synopsys tools to confirm the block still behaves the same and to measure power, performance and area. OpenAI is now training GPT-Synopsys to use EDA tools the way expert designers do. Same pattern as Lean for math or test suites for code: the model can be creative because something mechanical says yes or no.
OpenAI also published a cleaner benchmark: two open RISC-V cores, PicoRV32 and Ibex, 9,881 scored runs, designs zeroed if they failed simulation or used more than 16 times the original area. At about 112,000 tokens per run GPT-6.2 Astra scored highest on both, roughly twice GPT-6 Astra on PicoRV32. Every model did better designing from a written spec than improving the existing core, and OpenAI flags why that may flatter them: spec designs can drop features like performance counters, and timing wasn't checked after physical synthesis.
Caveat: apart from AMD's own verification statements, which OpenAI relays, these are vendor numbers. Still, "AMD shipped some of it" is a harder claim than a leaderboard, and it closes the loop the industry keeps talking about. The models are now cheaper to run partly because the models helped make the chips cheaper.
← Back to all articles
Start with OpenAI's own inference chip, JalapeΓ±o. The team began using GPT-4o on the design in early 2025, and by OpenAI's estimate model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, silicon-proven at bring-up. One multiplier redesign used about 15% less area and dropped latency from two clock cycles to one. Then AMD: OpenAI reports up to 14% PPA improvement over heavily optimized baselines on AMD's flagship GPU IP, which it calls roughly a year of progress in under five months. AMD has put some of the changes into a design already sent for manufacturing, and AMD itself verified single-digit PPA gains on key blocks of its next flagship CPU.
The part worth copying is not the model but the referee. Every SystemVerilog rewrite was checked with Synopsys tools to confirm the block still behaves the same and to measure power, performance and area. OpenAI is now training GPT-Synopsys to use EDA tools the way expert designers do. Same pattern as Lean for math or test suites for code: the model can be creative because something mechanical says yes or no.
OpenAI also published a cleaner benchmark: two open RISC-V cores, PicoRV32 and Ibex, 9,881 scored runs, designs zeroed if they failed simulation or used more than 16 times the original area. At about 112,000 tokens per run GPT-6.2 Astra scored highest on both, roughly twice GPT-6 Astra on PicoRV32. Every model did better designing from a written spec than improving the existing core, and OpenAI flags why that may flatter them: spec designs can drop features like performance counters, and timing wasn't checked after physical synthesis.
Caveat: apart from AMD's own verification statements, which OpenAI relays, these are vendor numbers. Still, "AMD shipped some of it" is a harder claim than a leaderboard, and it closes the loop the industry keeps talking about. The models are now cheaper to run partly because the models helped make the chips cheaper.
Comments