How OpenAI Designed Its Jalapeño Chip With Its Own LLMs
OpenAI’s first AI accelerator, Jalapeño, went from concept to silicon in under 20 months, with the company’s own language models speeding up the design work.
What happened
IEEE Spectrum reports that Jalapeño, OpenAI’s debut AI accelerator, was designed with help from the company’s own large language models. The chip went from first architecture concept to first silicon in under 20 months, and only nine months separated the first RTL from tape-out.
Jalapeño delivers up to 13.4 petaflops of 4-bit compute, 232 GB of HBM4 memory and 15.4 TB/s of bandwidth. OpenAI says it can cut end-to-end latency by up to 3.6 times versus Nvidia’s GB300 while drawing less power.
Why it matters
“The models are giving superpowers to our engineers,” says Richard Ho, OpenAI’s VP of hardware. A team of fewer than 100 people worked alongside Broadcom, which handled physical design from the gates onward. Experts call the timeline credible — and expect it to look slow soon.