OpenAI's In-House Chip "Jalapeño" Debuts, Per-Watt Throughput 1.7x That of GB300
On August 26, OpenAI's in-house inference chip, Jalapeño, previously reported, saw its performance data publicly released for the first time. In tests across three models—GPT-OSS 120B, DeepSeek R1…
On August 26, OpenAI's in-house inference chip, Jalapeño, previously reported, saw its performance data publicly released for the first time. In tests across three models—GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T—Jalapeño's per-watt AI throughput was 1.5 to 1.9 times higher than that of GB200 and GB300, with per-watt throughput under DeepSeek R1 load reaching 1.7 times that of GB300. End-to-end latency was approximately 28% to 59% of that of the comparison systems.
[TechFlow]
Original: https://www.techflowpost.com/newsletter/detail_133528.html
insigtX content is informational and educational, not investment advice.