According to @OpenAI, Jalapeño delivers higher throughput and lower latency per watt in one architecture without sacrificing efficiency.