AI 导读
在 DeepSeek R1 670B 上，Jalapeño 改变了整个推理前沿：在测试速度范围内每瓦吞吐量更高，在效率相当的情况下延迟比 GB300 低约 4.9 倍。 随着工作负载从以效率为导向的服务转向高度交互式推理，Jalapeño 在整个曲线上始终保持领先。
正文
On DeepSeek R1 670B, Jalapeño shifts the entire inference frontier: more throughput per watt across the tested speed range, with ~4.9x lower latency at comparable efficiency than GB300.
Jalapeño stays ahead across the whole-curve, as the workload moves from efficiency-focused serving toward highly interactive inference.
来源：@rohanpaul_ai · x.com