Source item
Open original source ↗Broadcom is enabling OpenAI’s impressive Jalapeño chip.
Quoted context
These are the quoted author’s words.
Interesting look at how @OpenAI's first custom inference chip could change inference hardware competition, evolving beyond raw performance. On GPT-OSS-120B, Jalapeño appears to sit on the throughput-per-kW frontier across a wide range of decoding speeds, while delivering ~4.2x lower end-to-end latency at comparable efficiency versus GB200 STP. As inference scales, performance per watt and latency increasingly could matter as much as raw compute. The AI hardware stack is still far from settled. This is extremely impressive from @OpenAI.
MattarARK ↗