Jalapeño beats Blackwell on perf/watt with two caveats: HBM4 vs Rubin, no long-context tests
nrehiew_ · x · 2026-10-11
nrehiew flags two caveats to Jalapeño's headline perf/power win over Blackwell: it uses HBM4 (Rubin is the fairer comparison) and the results weren't measured at super long context. Notably, Jalapeño's numbers don't use spec decoding or MTP — effectively a handicap, making the result more impressive.
Related event: Researcher's notes unpack OpenAI's rumored Jalapeño chip design tradeoffs(8 posts)→
More from Infra
- 5 LLM deployment patterns every AI engineer should know, from API calls to hybrid — goyalshaliniuk · 2026-10-11
- Neoclouds that just buy power and GPUs will lose to software-first players, says Beam founder — edgarpavlovsky · 2026-10-11
- M5 Ultra 256GB vs dual DGX Spark: local LLM buyer weighs throughput vs memory — MasterNomie · 2026-10-11
- AI Infrastructure Borrowing Slumps 80% in Three Months, From $113B to $23B — TansuYegen · 2026-10-11
- OpenAI, Microsoft, Google and 4 others signed a pledge to cover AI data centers' grid costs — ChrisGPT · 2026-10-11
- One-town monopolies: Yiwu makes 80% of Christmas decor, every EUV machine comes from Veldhoven — sahilypatel · 2026-10-11