NVIDIA Shruti Koparkar 谈 AI 工厂能效经济与 CUDA 生态
NVIDIA 的人聊怎么把每瓦电榨出最多算力，还讲了 Pinterest 一套 GPU 换三种模型跑，搞基础设施的可以看看。
NVIDIA 的 Shruti Koparkar 在访谈中解释了 AI 基础设施的能效经济，她会优先测试每瓦性能（performance per watt）。她说明了 CUDA 与 CUDA-X 在 AI 工厂软件栈中的实际作用。访谈还提到 Pinterest 用同一套 GPU 集群先后跑推荐系统、语言模型和多模态视觉语言模型。她讲了新模型发布时的负载变化，以及何时保留、改造或替换旧 GPU 的判断标准。
How much useful work can an AI factory get out of the power it has?
I sat down with @nvidia 's Shruti Koparkar to talk about the economics and software behind AI infrastructure.
We discuss why she would test performance per watt first, what CUDA and CUDA-X actually do, and how Pinterest has used the same GPU fleet for recommendation systems, language models and vision-language models.
We also get into what happens when a new AI model launches, and when it makes sense to keep, repurpose or replace older GPUs.
YouTube down below:
