fast-math is a set of compiler flags that enables a series of optimizations for floating-point arithmetic.

• -ffast-math (and included by -Ofast) in GCC and Clang
• -fp-model=fast (the default) in ICC
• /fp:fast in MSVC
• --math-mode=fast command line option or @fastmath macro in Julia.

But please notice that fast-math is not free lunch. Use it with cautions 😲.

Hence check this post ¹ to see what innocuous performance speedup or unfortunate downstream effects fast-math would bring to you.

I mean, the whole point of fast-math is trading off speed with correctness. If fast-math was to give always the correct results, it wouldn’t be fast-math, it would be the standard way of doing math. -- Mosè Giordano ²

Today, let's enjoy anotehr (previous one is here) amazing interactive visualization of CRDTs¹.

An excellent visual article by the Financial Times on how generative AI and the transformer model work. Entry level, but very well done.

• Generative AI exists because of the transformer
• 中文翻译: Transformer 的诞生让生成式 AI 成为可能

We're excited to announce that the CUDA C++ Core Libraries (CCCL) - Thrust, CUB, and libcudacxx - are now unified under the nvidia/cccl repository.

-- Jake Hemstad (@jrhemstad) on GitHub¹

• thrust: The C++ parallel algorithms library.
• cub: Cooperative primitives for CUDA C++.
• libcudacxx: The C++ Standard Library for your entire system.

I delightful love news / work like unifying abstractions. It makes me peace, relaxy and makes everything robuster.

Refs:

• NVIDIA/cccl: CUDA C++ Core Libraries
• Bundle CCCL in CuPy #7851

Today I got to update my #ObscureKnowledge that the system epoch (1970-01-01) is not the same as the @PostgreSQL epoch (2000-01-01) to help someone squash a bug.