跳到主要内容Skip to main content

CUDACUDA

CUDA

  • 硬件与部署Hardware & deployment
  • 第五站 · 跑起来Station 5 · Running It
  • 引用核验 2026-11-17refs checked 2026-11-17
解决什么问题What problem it solves

NVIDIA GPU 的并行计算平台,也是多数训练框架默认的底层执行环境。NVIDIA’s parallel computing platform and the default execution layer for most training frameworks.

这是什么What it is

CUDA 提供编程模型、编译器与库,让开发者把计算任务交给 GPU 的数千个核心。

PyTorch 等框架的 GPU 加速建立在 CUDA 及其库(如 cuBLAS、cuDNN)之上。

CUDA provides a programming model, compiler and libraries so work can be spread across thousands of GPU cores.

PyTorch’s GPU acceleration builds on CUDA and libraries such as cuBLAS and cuDNN.

CUDA is the parallel execution layer on NVIDIA GPUs: frameworks hand operators to the CUDA runtime, and the driver schedules them across thousands of cores.
CUDA 是 GPU 上的并行执行层:框架把算子交给 CUDA 运行时,再由驱动调度到数千个核心。CUDA is the parallel execution layer on NVIDIA GPUs: frameworks hand operators to the CUDA runtime, and the driver schedules them across thousands of cores.
为什么需要知道Why it matters

它解释了为什么「有 GPU」不等于「能跑」:驱动、运行时、框架版本要匹配;也是 CUDA 生态锁定讨论的起点。It explains why having a GPU is not the same as being able to run: driver, runtime and framework versions must match. It is also the starting point of the CUDA lock-in debate.

别名Chinese aliases

统一计算架构 · NVIDIA CUDA

出处Sources
  • NVIDIA · CUDA ToolkitNVIDIA CUDA Toolkit ↗
  • PyTorch · CUDA 语义说明PyTorch CUDA semantics ↗