Why are "tiny" types like std::float8_t and std::int4_t being proposed for C++29?

The question answers itself before it can be asked.

From P4025R0:

2. Ultra-Low Precision Arithmetic for modern LLMs & Quantization

   • Goal: C++23 has float15t and bfloat15t. But we need native support for std::float8t (FP8) and std::int4t for TPUs, and "nibble" types for weight compression.
   • Need: Without standard ultra low-precision types, std::linalg and std::simd cannot be used for modern AI inference. We also need a Quantization Schema for mdspan that automatically handles the de-quantization math such as scale and zero_point logic on access.

So yes, looks like it is for LLMs weight compression (as well as other TPU related needs).

It is mostly for SIMD: single instructions, multiple data. These are used (small floats) on for 3d rendering (colours: brightness can have several order of magnitude, but we may care less on exact colours (so significant bits).

I'm also pretty sure this is the wrong question type. There is a definitive why in the paper that proposes the change. What advice is there to be had about the matter?