Update on Unums
Nick B via Digitalmars-d-announce
digitalmars-d-announce at puremagic.com
Tue Mar 14 14:26:22 PDT 2017
On Tuesday, 14 March 2017 at 19:32:57 UTC, H. S. Teoh wrote:
> On Tue, Mar 14, 2017 at 06:50:07PM +0000, Nick B via
> Digitalmars-d-announce wrote:
>> On Tuesday, 14 March 2017 at 13:38:09 UTC, jmh530 wrote:
>> > On Tuesday, 14 March 2017 at 08:21:03 UTC, Andrea Fontana
>> > wrote:
>> > >
>> > > It seems public:
>> > > http://insidehpc.com/2017/02/john-gustafson-presents-beyond-floating-point-next-generation-computer-arithmetic/
>> >
>> > Also in pdf here
>> > http://web.stanford.edu/class/ee380/Abstracts/170201-slides.pdf
>>
>> Thank you both for posting these links :).
> [...]
>
> Indeed.
>
> But while the .pdf mentions Posits and Valids, the following
> slides only discuss Posits. Where's the discussion on Valids?
>
> In spite of that, though, Posits appear to be a much better
> candidate at replacing IEEE 794 floats than the previous unum
> incarnations. I felt the previous incarnations, while clever
> and workable in theory, posed too many practical challenges to
> implement on silicon. The current description of Posits seem
> to be much more feasible to put on silicon.
>
> Still, though, I wonder what Gustafson has up his sleeves wrt.
> Valids.
In the Stanford presentation pdf, (note that these change from
presentation to presentation) on page 12, is the only mention of
Valids. I thought these were a rename of Sets of Real
Numbers(SORNS) from his Type 2 Unums, but after reviewing the
slides again, I'm not sure. I believe that this needs to be
vertified with Dr Gustafson, as you correctly point out there are
no examples of Valids.
But when I review his slides from his New Zealand talk, there is
an additional slide, where he states that (1) "Posit pairs beat
intervals at their own games, too: Valid mode." and (2) "Posit
mode: Round unum after every operation. Valid mode: rigorous
answer bounds; NaN answers are sets. "
If anyone wants a copy of these New Zealand slides, please advise
me of your email address.
More information about the Digitalmars-d-announce
mailing list