AI 精选动态
智能评分 80
NVIDIA Blackwell 推理栈一月内将 DeepSeek V4 token 成本削减 5 倍
AI 推荐理由
该报告首次公开 5 倍的成本降低，为大规模部署提供了重要经济依据。核心解读
NVIDIA 在最新报告中指出，其 Blackwell 推理栈在一个月内使 DeepSeek V4 的 token 成本降低至原来的一五分之一。报告强调该成本优势来自于对 LLM、CPU、DPU、CUDA、网络、内存、安全和工具的统一协调。
全文
NVIDIA's newly published report says its Blackwell inference stack cut DeepSeek V4 token costs by up to 5x in one month. https://t.co/hoEquQQ3zW
Rohan Paul (@rohanpaul_ai): Agents stretch one prompt across models, tools, memory, GPUs, CPUs, DPUs, and storage.
NVIDIA’s cost edge comes from coordinating LLMs, CPUs, DPUs, CUDA, networking, memory, security, and tools. https://t.co/JCOwMBgUO6
Rohan Paul (@rohanpaul_ai): https://t.co/bxuYeVms34 https://t.co/5SUrOh0z4j