OpenAI and Broadcom announce chip designed for LLM inference at scale
Summary
Ars Technica reports OpenAI and Broadcom unveiled Jalapeño, a custom ASIC designed for large language model inference in data centers. The chip aims to improve performance per watt and reduce reliance on external GPUs, with deployment planned by year-end and a roadmap for future generations. Development reportedly took nine months and incorporates OpenAI’s model roadmap insights.