AI Hardware
OpenAI and Broadcom unveil Jalapeño chip for large-scale LLM inference
The specialized ASIC, developed over nine months, promises better performance-per-watt efficiency than existing data center hardware for language model workloads.
Ars Technica
OpenAI and chip manufacturer Broadcom have unveiled Jalapeño, a custom silicon designed from the ground up for running large language model inference at scale in data centers.
The ASIC was developed over nine months in close collaboration with OpenAI's research team, drawing on the company's insights into future model requirements and product roadmaps. Unlike general-purpose processors currently used for LLM inference, Jalapeño targets the specific computational patterns of modern language models.
What's being claimed:
The partners assert that early benchmarks show the chip delivers "substantially better" performance-per-watt compared to existing state-of-the-art alternatives. However, OpenAI acknowledged that full performance validation is still underway, with a technical report expected in coming months.
This marks the first generation of what both companies describe as a long-term effort to iteratively refine custom chips tailored to evolving LLM workloads—signaling OpenAI's deepening commitment to controlling its hardware infrastructure alongside its software innovation.
