@dougmerritt @kenshirriff The only thing that is due to the already progressing work on the 8087 is the extended precision in IEEE 754, built such that the 80 bit thingy is included. At least that is what I was told.
@jpmzemke
You're saying that you were told that the non-extended precision 8087 was complely independent?
But that's already contradicted by my quote above; the project started in 1977 and the co-architect Palmer credited Kahan, and Kahan was 90% or more of IEEE 754.
Clearly they did a ton of work way before IEEE was finalized, so it's not that Palmer et al just copied IEEE, which didn't exist yet, but I'm trying to clarify that it's *also* true that IEEE wasn't a merely a copy of the 8087.
Kahan is called "The Father of Floating Point" for strong reasons.
Deeper more nuanced questions of credit splitting I have no idea about, but that *is* about nuance.
I'm sure the architects of the 8087 did lots of work, I'm not saying otherwise.
Floating point is a huge subject.
P.S. Numerical analysts and floating point experts say that, if you have accuracy problems, you can't just blindly go from 64 bit to 80 bit to fix them -- that's a fool's paradise.
80 bits helps as part of a solution, not as the whole solution.
But I digress.
Oct 18, 2026, 6:07 PMYes.
I think the point is that the IEEE committee added the parts about _extended_ formats to the standard at least in part to reflect Intel's work on the 80-bit format for the 8087.
If I recall correctly, such extended formats are separate from the "main sequence" of 32-64-128-bit formats.
Again, this is all from memory, so don't take my word for it.
@vnikolov @dougmerritt @kenshirriff That is how I learned it. My memory says that I got that from my former supervisor Siegfried Rump, but it might have been another person, even Beresford Parlett.
@jpmzemke
Parlett was a high powered guy. I see we just lost him in February at age 94 (!). At Berkeley he was chair of C.S., and "He was the author of many influential papers on the numerical solution of eigenvalue problems, the QR algorithm, the Lanczos algorithm, symmetric indefinite systems, and sparse matrix computations.[3]"
https://en.wikipedia.org/wiki/Beresford_Parlett
Rump is head of the Institute for Reliable Computing (great topic).
> What is research?
> Yesterday I finished a proof,
> Today I found a flaw.
> I had a pleasant night.
https://www.tuhh.de/ti3/rump/
The history is clearly complicated and interesting, and I certainly don't know all of it.
Back in the day, even before IEEE 754 was implemented, I was following the work of Kahan, whose research and writings I admired, and I worked with a (later, less famous) head of the 754 committee, Dick Delp, which gives me one perspective, but I claim no expertise; I just have a long interest.
Come to think of it, I should clarify that I have even *less* expertise than may be implied by the above. And interest is certainly never the same as knowledge.
Oct 18, 2026, 6:49 PMI appreciate this information.
Needless to say, Kahan's work is worth reading indeed.
Just in case it isn't redundant, this is another good paper:
David Goldberg
What Every Computer Scientist Should Know About Floating-Point Arithmetic
(1991)
One place is
<https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html>.
@vnikolov
> David Goldberg
> What Every Computer Scientist Should Know About Floating-Point Arithmetic (1991)
Not only not redundant, it's a "must read" classic.
What it says is *still* not widely enough known.
Thanks for the link.
Oct 19, 2026, 1:08 AM@dougmerritt @vnikolov @jpmzemke @kenshirriff Based on the common error patterns, very clearly a lot of the folks building frontier machine learning systems haven’t read it.
@dougmerritt @kenshirriff No, not independent. But the 80 bit had already been planned and were added for that reason to IEEE 754. 32, 64, and 128 bit feel more natural, why should one have a format that has more bits than 64 and not directly quadruple precision?
@jpmzemke
Intel has made a lot of strange decisions over the decades. I won't defend 80 bits, but if I were to hazard a guess, perhaps it fit better into their transistor budget that year than did 128 bits. There may have been added concerns about the number of cycles burned.