AI Signal 142
OpenAI's Jalapeño inference ASIC beats Nvidia Blackwell on perf/W in SemiAnalysis lab benchmarks but still in engineering samples
↗
OpenAI announced Jalapeño, a custom inference ASIC built with Broadcom, which SemiAnalysis benchmarked in OpenAI's labs and found beating Nvidia Blackwell on performance per watt across multiple open-source models.
The chip was designed from scratch to tapeout in roughly 16 months and uses HBM4, making it a closer competitor to Nvidia's Rubin than to the currently shipping Blackwell. All benchmark numbers were provided by OpenAI and the full benchmark suite was not run, so the results are preliminary.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
OpenAI's Jalapeño ASIC was designed in partnership with Broadcom, going from team hiring to tapeout in approximately 16 months.
SemiAnalysis verified InferenceX benchmark runs in OpenAI's lab, finding Jalapeño beats Blackwell on perf/W without multi-token prediction or speculative decoding.
The comparison is incomplete because Jalapeño is still engineering samples while Nvidia's Rubin, which also uses HBM4, is already shipping to customers.
THE CLUSTER