Choosing the right news sources is important. In my case Hacker News 50 proved to be precious as in this recent post/toot/message they proposed “Fifteen Years of FP64 Segmentation, and Why the Blackwell Ultra Breaks the Pattern“.
Briefly, modern consumer GPU has a 64:1 ration in FP64:FP32 performances (FP64 is a shortcut for Double-precision floating-point number format).
Now I do want to implement – in Eiffel ça va sans dire – T. J. DEKKER’s Technique for Extending the Available Precision in Floating Point aritmetics