OpenAI用自有大语言模型设计首款AI加速器芯片

OpenAI于8月25日推出了首款自研AI加速器芯片Jalapeño1。该芯片从初始架构到首版硅设计仅用时不足20个月1，其中从首个RTL设计到流片耗时9个月1。Jalapeño交付13.4 petaflops的4位计算能力，配备232GB内存和15.4 terabytes/秒带宽1。与Nvidia GB300相比，Jalapeño在功耗更低的情况下可将端到端延迟降低3.6倍1。

OpenAI利用自有大语言模型加速了芯片设计流程，特别是在高级综合和软件优化阶段1。设计团队规模平均不足100人，涵盖硬件、软件和供应链等多个角色1。OpenAI与Broadcom合作进行该项目，OpenAI负责系统设计，Broadcom负责物理设计1。Jalapeño项目始于2024年10月，团队使用了包括o3和GPT-6 Astra前身在内的模型1，以及Google开源的高级综合工具XLS和经过内部微调的芯片设计专用大语言模型1。AI辅助物理设计优化实现了矩阵乘法单元面积减少10%1。

首批芯片回厂后约40小时内，DeepSeek多头潜在注意力内核基准性能从0.31%提升至88.94%1。OpenAI硬件副总裁Richard Ho表示"模型为工程师提供了超级能力"1，技术人员Chris Leary指出"AI在软件类任务上表现更好"1。

OpenAI introduced its first AI accelerator chip, Jalapeño, on August 25th, completing the journey from initial architecture to first silicon in under 20 months 1. The company deployed its own large language models to streamline the chip design process, enabling a team of roughly 100 engineers across hardware, software, and supply chain roles to iterate rapidly through advanced synthesis and software optimization phases 1.

The Jalapeño chip delivers 13.4 petaflops of 4-bit computing power alongside 232GB of memory and 15.4 terabytes per second of bandwidth 1. In performance comparisons, it reduces end-to-end latency by 3.6 times relative to Nvidia's GB300 while consuming less power 1. The design progression from initial register-transfer language to tape-out took just 9 months 1. OpenAI partnered with Broadcom, with OpenAI handling system design while Broadcom managed physical design implementation 1.

The development process, which began in October 2024, incorporated multiple AI models including o3 and predecessors of GPT-6 Astra 1. The team utilized XLS, Google's open-source high-level synthesis tool, alongside internally customized large language models specifically tuned for chip design tasks 1. AI-assisted physical design optimization achieved a 10 percent reduction in matrix multiplication unit area 1. Following the first chips' return from fabrication, performance on DeepSeek's multi-head attention kernel benchmark improved from 0.31 percent to 88.94 percent within approximately 40 hours 1. Richard Ho, OpenAI's vice president of hardware, stated that "models provided engineers with superpowers," while technical contributor Chris Leary noted that "AI performed better on software-type tasks" 1.