案例库 · 软件与 IT · 战略决策 · 2006–2020
这条还没译成中文，下面是英文原文。
CUDA made the gaming GPU a computing standard by giving it a free toolchain.
Nvidia spent years making its GPU programmable with a free C-style compiler and a developer ecosystem, so when AI came, CUDA was the only mature standard.
NVIDIA
那一手
By the mid-2000s a graphics chip had hundreds of cores that could do far more than shade triangles, but using them meant writing graphics code. In November 2006 Nvidia unveiled CUDA and the industry's first C-compiler for the GPU, letting ordinary programs run on the parallel hardware.
CUDA was opened to developers and libraries, so researchers and companies could use it without Nvidia writing their software. The company invested heavily and went years without turning a profit on the platform, betting that general-purpose GPU computing would eventually matter.
That bet landed with deep learning, which is exactly the kind of massively parallel math a GPU does well. Because CUDA and its ecosystem were already mature and free to use, it became the default platform for AI frameworks and training, and rivals had to catch up.
为什么管用
- The free C-compiler and toolkit let any programmer use the GPU without learning graphics.
- Years of building libraries and developer tools created a deep, hard-to-replace ecosystem.
- A long investment meant CUDA was mature long before AI demand appeared.
- Every major AI framework was built on CUDA, so its installed base kept attracting more developers.
可以搬走什么
A standard takes hold when the enabling part is free and the developer base is built before the demand arrives, so the technology is ready exactly when the opportunity does.
后来呢
CUDA became the de facto standard for AI and scientific computing, and Nvidia's data-centre business grew with it. Alternatives exist, but the decades of developer lock-in have made CUDA the default that new entrants must match.
资料来源
发现哪里写错了？告诉我们。