OpenAI has announced the test results of "Jalapeño," its first custom-developed inference ASIC. According to the company, the chip achieves both maximum throughput and high response speeds (low latency) within a single architecture—a balance that the company stated had been difficult to achieve with existing hardware systems.

In tests using SemiAnalysis's "InferenceX" benchmark, OpenAI reported that for public models including GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, AI processing per watt at peak throughput increased by 1.5 to 1.9 times, while end-to-end latency was reduced by 1.7 to 3.6 times. Additionally, the chip demonstrated performance improvements of 2.1 to 4.1 times for workloads requiring high interactivity.

Regarding the development process, OpenAI stated that by utilizing its own AI models and coding tools from the design phase, it completed the process from design initiation to tape-out in approximately nine months. OpenAI plans to begin integrating the chip into its computing infrastructure within 2026.

Source:

Source: