And is it only about the format or also about the behavior of the types under common mathematical operations?

And format (as I think was mentioned) can be storage format or range of mantissa and exponent.
 

-----Ursprüngliche Nachricht-----
Von: Paul Caprioli via Std-Proposals <std-proposals@lists.isocpp.org>
Gesendet: Sa 18.04.2026 22:21
Betreff: Re: [std-proposals] ISO/IEC 60559 Classification for std::numeric_limits
An: std-proposals@lists.isocpp.org;
CC: Paul Caprioli <paul@hpkfft.com>;

>> The AVX10.2 ISA (Intel/AMD) appears to include conversion instructions for
>> the quarter-precision types E5M2 and E4M3 described in the “OCP 8-bit
>> Floating Point Specification”.

> Are the mantissa and exponent counts sufficient? Or do we need to describe the
> exponent bias, whether the type has non-finite numbers, etc.?

What's in a name? That which we call a rose,
By any other word would smell as sweet.

In "Intel Advanced Vector Extensions 10.2 Architecture Specification" (document number 361050-005US), the instruction mnemonics for the two types use bf8 and hf8, respectively.

It makes sense to use the term bfloat8 for "E5M2". For example, the bit representation for 1.0 in IEEE Std. 754-2019 binary16 is:

S-E5-----F10----
Binary:0011110000000000

and 1.0 in bfloat8 is:

S-E5--F2
Binary:00111100

This is analogous to the relationship between IEEE binary32 and bfloat16.

As you hinted, the "E4M3" format does not have infinities (but it does have NaN).
Details are in the OCP specification, so I won't bore everyone on this list any further.

By the way, I appreciate that Shakespeare dereferenced his pointer. He wrote, "that which we call a rose", not merely "a rose". :)

 

-- 
 Std-Proposals mailing list
 Std-Proposals@lists.isocpp.org
 https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals