OpenAI Says Self-Developed Jalapeno Inference Chip Outperforms Nvidia GB300 in Tests, Plans Deployment This Year
- OpenAI's Jalapeno AI chip outperformed Nvidia's GB300 in tests.
- The Jalapeno chip leads in processing capability per unit of power and response speed.
- It consumes about 700 watts and is designed for AI inference, not model training.
- OpenAI plans to begin using the Jalapeno chip later this year.
- The chip has not yet been tested against Nvidia's next-generation Vera Rubin chip.
OpenAI announced that its self-developed Jalapeno AI chip, created in collaboration with Broadcom, has surpassed Nvidia's GB300 in performance tests. The Jalapeno chip excels in two key areas: processing capability per unit of power consumption and response speed. It operates at approximately 700 watts and is intended primarily for AI inference tasks rather than for model training.
The company plans to begin deploying the Jalapeno chip later this year. In public testing, OpenAI utilized one of its smaller open-source models and third-party models from DeepSeek and Moonshot AI, with the performance advantage being particularly notable with the Kimi model.
However, it is important to note that the Jalapeno chip has not yet been compared to Nvidia's next-generation Vera Rubin chip, which has recently started shipping. OpenAI also indicated that the second-generation chip is in the late stages of development and is expected to complete tape-out in the coming months, while reiterating that Nvidia remains a significant supplier of chips for the company.
OpenAI称自研的Jalapeno推理芯片在测试中优于Nvidia GB300，计划今年部署
OpenAI宣布，其与博通合作开发的自研Jalapeno AI芯片在性能测试中超越了Nvidia的GB300。Jalapeno芯片在每单位功耗的处理能力和响应速度两个关键领域表现优异。该芯片功耗约为700瓦，主要用于AI推理任务，而非模型训练。
该公司计划在今年晚些时候开始部署Jalapeno芯片。在公开测试中，OpenAI使用了其较小的开源模型以及来自DeepSeek和Moonshot AI的第三方模型，Kimi模型的性能优势尤为明显。
然而，需要注意的是，Jalapeno芯片尚未与Nvidia的下一代Vera Rubin芯片进行比较，该芯片最近已开始发货。OpenAI还表示，第二代芯片已进入开发后期，预计将在未来几个月内完成流片，同时重申Nvidia仍然是公司重要的芯片供应商。