OpenAI's Jalapeño chip reportedly beats Nvidia GB200 in benchmarks
mark_k · x · 2026-08-26
Reports from Hot Chips suggest that OpenAI's first inference chip, Jalapeño, performs better than expected. Benchmarks against models like GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T show it delivers 1.5-1.9x better performance per watt and 1.7-3.6x lower end-to-end latency than Nvidia's GB200/GB300 Blackwell systems. Notably, OpenAI used its own AI (Codex + GPT-Astra) to design the chip, with AI-generated kernels outperforming expert-written ones by 1.5-1.8x in specific blocks.
Related event: OpenAI's Custom Jalapeño Chip Beats Nvidia Flagships in First Tests(41 posts)→
More from Infra
- Snowflake reveals Agent hidden costs, reducing trial cost by 33%-45% — StasBekman · 2026-08-26
- M5 Ultra Studio pricing is wonderfully broken: 2-3x DGX Spark value for local AI — SumitGup · 2026-08-26
- Energy-first AI hardware design might mimic the brain — prateekj · 2026-08-26
- AC2 infrastructure helps enable post-trained specialist models — ypatil125 · 2026-08-26
- A beachside coal-powered data center, "water-cooled as God intended" — bennash · 2026-08-26
- Mac Studio vs RTX 6000: Local AI development cost and experience comparison — SumitGup · 2026-08-26