How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip
OpenAI's Jalapeño AI accelerator, co-designed with Broadcom, went from concept to silicon in under 20 months using LLM-driven design workflows.
OpenAI unveiled Jalapeño, its debut AI accelerator delivering up to 13.4 petaflops of 4-bit compute with 232GB of memory accessed at 15.4 TB/s, and citing up to 3.6x lower end-to-end latency than Nvidia's GB300. The chip moved from first architecture concept to first silicon in under 20 months, with only nine months between first RTL and tape-out, aided by internal LLMs and Google's open-source XLS high-level synthesis toolchain. A team averaging fewer than 100 people handled system design while Broadcom managed physical design from the gates onward. OpenAI's internal models also lifted a DeepSeek multi-head latent attention kernel from 0.31% to 88.94% of theoretical peak in roughly 40 hours.
- Jalapeño offers 13.4 petaflops of 4-bit compute and 232GB memory at 15.4 TB/s
- OpenAI benchmarks claim up to 3.6x latency improvement over Nvidia's GB300
- LLM-assisted design plus XLS cut RTL-to-tape-out to nine months
- Team of fewer than 100 engineers partnered with Broadcom on physical design
- Internal models optimized an inference kernel to 88.94% of theoretical ceiling
- Products
- Jalapeño
- Organizations
- UC San DiegoUniversity of Maryland
From the original article
7 min read at spectrum.ieee.org
On 25 August, OpenAI fully unveiled Jalapeño, the company’s debut AI accelerator chip. Jalapeño delivers up to 13. 4 petaflops of 4-bit compute and accesses 232 gigabytes of the most advanced memory available, linking to it at a blazing 15.
By maxall4 · spectrum.ieee.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.