Stanford Seminar: Beyond Floating Point: Next Generation Computer Arithmetic

Channel:Stanford Online · Length:1:31:28 · Published:2017-02-03 · Views:32864

Watch on YouTube

Summary

John Gustafson presents posits, a proposed drop-in replacement for IEEE 754 floats that is simpler in hardware and more flexible in precision. The seminar includes a live demo and machine learning results.

Why HN recommends it

Cited in floating-point threads as the main introduction to posits and unums as an alternative to IEEE floats.

Here's a nice introductory presentation — HN, 2018

Machine learning & deep learning · Pure mathematics · Computer architecture & CPUs

Subtitles (excerpt)

so I'm going to talk about probably the culmination of about 35 years of attempts to find something better than floating Point arithmetic starting in the 80s I think is when I started working on this right now I'm expatriated to University of Singapore and they' provided a very good home for me to work on this kind of research and I appreciate that so why worry about floating point I came up with a simple example because I found out that everybody trusts Matrix Matrix multiplication they use it for examples all the time well suppose just one row in one column where happen to be this a and that b so you were taking their scalar product in order to find out the result and these are fairly ordinary numbers these are integers they're not that big 32 million 1 negative 1 80 million 40 million no perfectly ordinary reasonable numbers if you do this in single Precision you'll get a. b equals z which would tell you hey these vectors are orthogonal might be an important property to know check it by doing double precision and you again get A.B equals 0 and you know what's coming uh let's suppose you do it with a binary sum collapse where you don't do it from left to right but you do it these two these two

On Hacker News

For anyone interested in this sort of thing you'd probably also be interested in John Gustafson's work on unums/posits. Here's a nice introductory presentation [VIDEO] — gbrown_, 2018-06
You may like: [VIDEO] It turns out the math is slightly harder but it's faster than IEEE FP in hardware, probably because of less conditionals in the spec. There is better performance in terms of numerical accuracy at the cost of substantially more difficult error analysis. Disclaimer: I'm in the video. — dnautics, 2017-09
Have you seen "Stanford Seminar: Beyond Floating Point: Next Generation Computer Arithmetic"? [VIDEO] — espeed, 2017-02
Well unums in particular (especially the new type 3 "posits" ones) are very interesting and practical for hardware implementation. John Gustafson, the creator of unums, gave the first talk on type 3 unums at Stanford earlier this month, and there are already multiple hardware implementation efforts (one of which I am indirectly overseeing). While many people have recently hyped up low precision f… — trsohmers, 2017-02
I'm very surprised nobody has brought up John Gustafson's Unum work. His latest iteration is the posit, which he presented at Stanford earlier this month[0][1] EDIT: apologies to CalChris[2]. What I should have said is discussed Gustafson's Unum work. [0] [VIDEO] (relevant question at 1h22m37s) [1] [LINK] [2] [LINK] — vanderZwan, 2017-02

Related videos