📄 中文摘要
2026年，运行大型语言模型的关键不再是是否可以在本地完成，而是与云服务提供商相比，其财务和性能是否合理。该基准测试全面比较了NVIDIA DGX Spark在本地LLM推理性能与主要云服务提供商的表现，提供了真实的数据，帮助用户做出明智的决策。测试方法包括硬件配置，使用了GB10 Grace Blackwell超级芯片、128GB统一LPDDR5x内存、2TB NVMe SSD存储和Ubuntu 22.04 LTS操作系统，结合CUDA等软件工具，确保了测试的准确性和可靠性。
DGX Spark 推理性能：本地 LLM 与云基准比较（2026）
出处: DGX Spark Inference Performance: Local LLM vs Cloud Benchmarks (2026)
发布: 2026年3月19日
📄 English Summary
DGX Spark Inference Performance: Local LLM vs Cloud Benchmarks (2026)
In 2026, the focus shifts from whether large language models can be run locally to whether doing so is financially and performance-wise advantageous compared to cloud providers. This benchmark provides a comprehensive comparison of NVIDIA DGX Spark's local LLM inference performance against major cloud providers, offering real-world data to aid in informed decision-making. The testing methodology includes hardware configuration featuring the GB10 Grace Blackwell Superchip, 128 GB unified LPDDR5x memory, 2TB NVMe SSD storage, and Ubuntu 22.04 LTS operating system, along with software tools like CUDA to ensure accuracy and reliability in testing.