OpenAI says models cut Jalapeño area and improved AMD GPU IP
First merged summary.
OpenAI says its models produced silicon-proven Jalapeño changes and up to 14% PPA gains on AMD GPU IP, with some changes already manufactured.
OpenAI says successive models, beginning with GPT-4o in early 2025, proposed power, performance and area changes to its Jalapeño inference chip that were silicon-proven at bring-up. It reports about 10–15% cell-area cuts on Jalapeño blocks; one multiplier redesign cut cell area about 15% and latency from two cycles to one. With AMD, OpenAI reports up to 14% PPA improvement versus optimized flagship GPU IP baselines in less than five months, which it equates to about a year of earlier engineering, plus single-digit PPA gains on an upcoming flagship CPU. AMD has included some model-written changes in a design already sent for manufacturing. OpenAI scored model generations on PicoRV32 and Ibex, where GPT-6.2 Astra led, and is working with Synopsys on GPT-Synopsys to connect models to existing EDA tools.
- GPT-4o began assisting Jalapeño design in early 2025.
- A multiplier redesign cut cell area about 15% and latency from two cycles to one.
- OpenAI reports up to 14% PPA gain on AMD flagship GPU IP in under five months.
- Single-digit PPA gains reported on an upcoming AMD flagship CPU.
- GPT-6.2 Astra scored highest on PicoRV32 and Ibex at about 112,000 tokens.
- OpenAI is building GPT-Synopsys with Synopsys for AI-assisted EDA.
What each outlet adds
beyond the shared facts- OpenAI NewsJalapeño cell-area cuts, CPU gains, and PicoRV32 and Ibex scores.
- The DecoderUnder five months of work equated to about a year of AMD engineering; names GPT-Synopsys.
Coverage
oldest first · one row per article · UTCOpenAI models used to improve Jalapeño and AMD chip designs
OpenAI reports about 10–15% cell-area cuts on Jalapeño blocks and up to 14% PPA gains on AMD GPU IP.
OpenAI says models added a year of AMD GPU optimization
OpenAI claims up to 14 percent PPA gain on AMD flagship GPU IP, with some changes already sent to manufacturing.