OpenAI Announces Performance Results for Its First Custom Chip, Jalapeño
Performance results for Jalapeño, the first custom chip developed in the artificial intelligence field, have been shared on the InferenceX test.
OpenAI has shared the initial measured performance results for its first custom chip, Jalapeño. Evaluated in the publicly known InferenceX tests, the chip offered higher efficiency per kilowatt consumed and lower latency times.
InferenceX Test Results
The Jalapeño chip developed by OpenAI was successfully evaluated in the public InferenceX benchmark test created by SemiAnalysis.
During the tests, the capacity of the chip system was measured using different model families such as GPT-OSS 120B, DeepSeek R1, and Kimi K2.5.
Efficiency and Speed Advantages
In the measurements performed, the Jalapeño chip performed significantly more AI operations per watt at peak efficiency compared to commercial systems.
In addition, marked improvements were achieved in token latency times and end-to-end latencies compared to standard hardware.
Power Consumption Values
It was stated that the Jalapeño chip is architecturally rated and designed for 700 watts of power.
During the tests conducted, it was observed that the chip's sustainable power consumption remained at or below the 550-watt level.
Future Plans
The Jalapeño hardware is planned to begin integration into the company's own infrastructure by the end of the year.
This chip represents the first step of a multi-generational roadmap, and work is ongoing for subsequent new generations.