Fortran 70年仍统治科学计算
推荐指数 67.0 NO. 014 · 2026.10.16
发布2026/10/15Score175Comments75
为什么值得看
Fortran 发布首版手册70周年,至今仍是核工程、气候预测、材料科学等领域的首选语言。对 AI 工程师的启示:科学计算基础设施的迁移成本极高,Python 生态看似繁荣,但底层数值计算仍大量依赖 Fortran 库(如 NumPy 调用 LAPACK)。
编辑判断
AI 圈很少讨论的一个事实是:你每次调用 NumPy 的矩阵运算,底层大概率在跑 60 年前就开始优化的 Fortran 代码。LAPACK、BLAS 这些库经过几十年硬件适配,性能调优到了极限,Python 和 C++ wrapper 只是包装层。
这意味着两个现实:第一,想做科学计算基础设施国产替代的团队,绕不开重写或兼容这些 Fortran 遗产;第二,大模型训练中的线性代数瓶颈,短期内在 Fortran 层优化的收益,可能高于上层框架的改动。剑桥这次庆典的真正价值,是把"老技术为何不死"的 engineering trade-off 摆上了台面。
社区反馈
正面 75 条评论
核心争论:Fortran 70年长寿是技术惯性还是真正不可替代
My first language, I should try and mess with it again. It has been decades since I used it.
I believe it has changed quite a lot, keeping its focus for scientific/engineering computing, but adapting more modern language constructs. I'm not sure how much of the archaic stuff (COMMON blocks, ENTRY points etc) is supported in the more recent standards for backward compatibility.
Aren't COMMON's implemented as static structures? Stuff that's accessible from multiple functions - except that Fortran requires explicit rather than implicit reference.