Hi y’all,

I see f16 mentioned in the spec, but not f8 (i.e. minifloats
<https://en.wikipedia.org/wiki/Minifloat>.) Is that intended?

At least Chibi Scheme supports f8 minifloats in numeric vectors:
https://github.com/ashinn/chibi-scheme/blob/8f5850c32d103dd6eeb4ffcd66e280e737f7e512/sexp.c#L3146
The choice of the exact format is queer: __NV_E5M2 from NVIDIA. This
highlights a problem though: everyone has their own favorite f8 format,
from the most conventional 1.4.3 (sign.exponent.mantissa,) to pretty
1.4.3-2 (my favorite,) to vendor-specific formats, to 1.2.1 (yes, f4!)

So yeah, I’m not particularly pushing this, as the choice of particular
f8 might be unpopular. But there’s a need for standardization here, as
there’s at least one implementation that already supports f<16.

Best of luck,
--
Artyom Bologov
https://aartaka.me