OpenAI announced performance benchmarks for Jalapeño, its new custom-designed AI chip developed in partnership with Broadcom. The company says the Application-Specific Integrated Circuit (ASIC) addresses a long-standing challenge in AI system design by simultaneously optimizing for both speed and throughput, metrics that typically require trade-offs according to OpenAI hardware vice president Richard Ho.
The chip is specifically engineered for AI inference, the computational process used to run trained AI models and deploy autonomous agents. To evaluate its capabilities, OpenAI used InferenceX, an industry benchmarking platform designed to measure inference performance. The testing compared Jalapeño against what OpenAI described as the best-performing alternatives available at the time of measurement, including Nvidia’s GB200 and GB300 superchips.
According to the company’s reported results, Jalapeño demonstrated substantial efficiency gains, delivering between 1.5 and 1.9 times more computational work per watt across three different AI models compared to the benchmark systems. Additionally, the chip achieved end-to-end latency reductions ranging from 1.7 to 3.6 times lower across the same models. Ho indicated these improvements would enable faster user responses, more responsive AI agents, and improved service reliability as demand increases.
OpenAI plans to begin limited deployment of Jalapeño by the end of this year, with intentions to significantly increase production volumes into 2027. The company has not disclosed specific manufacturing targets for the following year. Ho stated that Jalapeño will not replace OpenAI’s existing chip portfolio entirely, noting the company maintains partnerships with major vendors including Nvidia as part of its broader compute strategy. OpenAI indicated it intends to continue developing subsequent generations of the Jalapeño architecture.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI