OpenAI's Jalapeño inference chip beats Nvidia Blackwell in reported benchmarks

AI summary

OpenAI has publicly detailed Jalapeño, a custom ASIC built with Broadcom for LLM inference, after a tapeout process that began in mid-2024. According to benchmark results shared by the source, the chip outperforms Nvidia, AMD, and Google hardware on several inference metrics, especially throughput per watt, while remaining a general-purpose inference design rather than a model-specific accelerator.