OpenAI's 'Jalapeno' chip reportedly beats Nvidia Blackwell in efficiency and latency benchmarks
量子位 · wechat · 2026-08-26
SemiAnalysis benchmarks reveal that OpenAI's in-house inference chip 'Jalapeno' outperforms Nvidia's Blackwell in several metrics. With a TDP of just 700W, it achieves 1.5-1.9x the throughput per watt and 1.7-3.6x lower latency than Blackwell. Built on TSMC's N3P process with 216GB HBM4, the chip's design was heavily assisted by AI (e.g., Codex), with some AI kernels outperforming human experts. Engineering samples are ready, with mass production expected in 2027. Meanwhile, OpenAI's head of data centers Chris Malone has left the company amid executive turbulence.
Related event: OpenAI's In-House Chip Reportedly Beats Nvidia Blackwell on Efficiency(3 posts)→
More from Companies & People
- AI usage limits echo early billing eras; unlimited tokens ahead — RachelVT42 · 2026-08-26
- Jensen Huang: A decade of being ignored was Nvidia's compounding advantage — r0ck3t23 · 2026-08-26
- AI Coding Risks Preventing Expertise, Creating 'Expert Novice' Trap — addyosmani · 2026-08-26
- OpenAI Reportedly Developing Custom Chips, Codenamed Jalapeno — beffjezos · 2026-08-26
- OpenAI Accused of Unauthorized Annual Plan Upgrade, Discount Lost — phoenix177 · 2026-08-26
- OpenAI Architect: 'You Can Just Build Chips' — beffjezos · 2026-08-26