OpenAI公布自研推理芯片Jalapeño基准测试结果，性能超越当前最先进系统

OpenAI在Hot Chips会议上公布了其自研推理芯片Jalapeño的首批基准测试结果2。数据显示，在Semianalysis的InferenceX基准测试中，Jalapeño的每用户token数和每千瓦吞吐量均超越了包括Nvidia Blackwell系统在内的当前最先进推理处理器2。OpenAI发布博客称，该芯片在完成AI推理任务时比竞品更高效且响应更快1。

Jalapeño是一款专为AI推理设计的ASIC芯片，由OpenAI与Broadcom密切合作开发制造12。OpenAI硬件副总裁Richard Ho表示，Jalapeño能够提供更低延迟和更高吞吐量1。他进一步强调：“关键在于，结果显示其性能较现有技术有非常非常显著的进步。”2 同时他指出：“Jalapeño每单位功耗可处理更多AI工作，同时更快返回响应。服务大量客户非常高效，且能保持极低延迟。”2 该芯片于去年10月首次公布2，并于今年6月首次推出1。

在部署计划方面，Jalapeño预计将于2026年底“以非常小的数量”进行部署，并在2027年进行更大规模部署2。OpenAI还计划将Jalapeño打造为多代平台，协同开发AI产品、模型、芯片和内存2。

OpenAI has revealed the initial benchmark results for its custom AI inference chip, Jalapeño, at the Hot Chips conference 2. The chip is an application-specific integrated circuit (ASIC) developed in close collaboration with Broadcom and is designed specifically for AI inference tasks 12. According to the InferenceX benchmark by Semianalysis, Jalapeño surpasses current state-of-the-art inference processors, including Nvidia Blackwell systems, in both tokens per user and throughput per kilowatt 2. A company blog post also noted that the chip completes tasks more efficiently and responds faster than competitors 1. OpenAI Vice President of Hardware Richard Ho 1 highlighted the chip's capabilities, stating, "The bottom line is that the results show a very, very significant performance advance over state of the art" 2. He further explained, "Jalapeño can serve more AI work per unit of power, while also returning responses more quickly. It's very efficient to serve a lot of customers, but it can also be very low latency" 2.

Jalapeño was first introduced in June 1 and initially announced in October of last year 2. OpenAI plans to deploy the chip in very small quantities by the end of 2026, followed by a larger-scale deployment in 2027 2. Furthermore, the company intends to develop Jalapeño into a multi-generational platform to co-develop AI products, models, chips, and memory 2.