On the 25th, at the US Hot Chips conference, OpenAI announced that "Jalapeno," an AI inference semiconductor co-developed with Broadcom, demonstrated performance surpassing Nvidia's current "GB300" product in testing.

According to Richard Ho, head of OpenAI's semiconductor division, comparative tests pitted Jalapeno against the GB300, which had shown the highest performance in public benchmarks. Jalapeno outperformed the GB300 in both AI compute per watt and response speed. Ho explained that Jalapeno delivers high performance at 700 watts of power consumption, contributing to reduced operating costs for data centers where power is a major expense.

According to reporting by SemiAnalysis, the outlet confirmed the Jalapeno chip at OpenAI's invitation and conducted benchmarks using the InferenceX suite in their lab. The tests reportedly showed performance exceeding chips from Nvidia, AMD, and Google across several top open-source models. Jalapeno features a general-purpose design that does not specialize in specific inference tasks but delivers high performance across a wide range of scenarios, characterized by hardware-software co-design.

Design work began in mid-2024 and was completed in approximately 16 months, from team formation to tape-out for manufacturing. OpenAI had announced in June of this year that the development was achieved in a record short period through its partnership with Broadcom, which began last year.

Jalapeno is designed for the AI inference stage and is not designed for training. No comparative tests have been conducted against Nvidia's next-generation semiconductor "Vera Rubin," which has just begun shipping. OpenAI plans to start using Jalapeno for its own AI models later this year, with Ho stating that they will decide which models to run on Jalapeno in the future.

Sources (integrating reports from two outlets):