Zhipu deploys ZCube, raising inference throughput 15% on the same GPUs
推翻二十年组网逻辑,智谱落地ZCube,让同样的GPU多干15%的活
Zhipu deployed ZCube in a thousand-GPU GLM-5.1 production inference cluster, replacing ROFT while keeping GPUs, software stack, and business code unchanged; throughput rose by over 15%, TTFT P99 fell 40.6%, and switch plus optical module costs dropped by one third.
Why it matters: Zhipu swapped in its own ZCube networking on a GLM-5.1 thousand-GPU inference cluster, raising throughput over 15% and cutting P99 first-token latency 40.6% without touching GPUs or business code.