NVidia Patent: Processing Dual Format Floating-point Data

Jawed

Legend
Have we seem this patent already?

http://patft1.uspto.gov/netacgi/nph-Parser?patentnumber=7075539


This floating-point notation is relatively data intensive. Therefore, passing and processing such data can be expensive, particularly, in high bandwidth applications, such as graphics processing. In graphics processing environments, the precision afforded by standard exponential notation (i.e., 32-bit or 64-bit) is frequently excessive. In other words, the precision afforded by standard exponential notation does not meaningfully improve the resultant graphics output.

There is a large embedded base of graphics computational systems that rely upon standard exponential notation. In addition, future graphics computational systems are designed to support standard exponential notation.

In view of the foregoing, it would be highly desirable to develop a technique that is operative in a standard exponential notation computation environment, but afforded enhanced data transport and computation operations through use of a non-standard exponential data format.

---


And, can I ask people to be careful about linking US patents? Using the style here, because I found this patent by clicking on a link in another thread, whose link no longer points to the patent intended :!: - instead it points directly to this patent. Nice when you find something interesting. Crap when you wanted to read the original patent.

Jawed
 
  • Like
Reactions: Geo
I noticed that patent a few days ago, and I conclude that it's completely weird. Why would you want to do FP32->FP16 conversion like that, except maybe for FP16 constants (and why couldn't that be done on the CPU, if it's just for constants?)

There's a much more interesting NVIDIA patent out there that I haven't seen linked around too much yet, or if it was, then it didn't focus on what I'm focusing on. But meh, it's not because something is in a patent that it'll get implemented, so it's highly interesting from a theorical pov, but has to be taken with a grain of salt.


Uttar
 
Uttar said:
There's a much more interesting NVIDIA patent out there that I haven't seen linked around too much yet
It would be interesting to know what it was, and what your take on it is:idea:
 
Well, this seems, on the surface, to be talking about something that has been in nVidia GPU's since the NV3x. Perhaps this is just a more optimal method of achieving partial precision?

Anyway, what does sound very interesting to me is the mention of double precision: does this mean that the G80 will have double precision ALU's? It didn't seem possible at all before, and even after this seems unlikely...but the hint is tantalizing!
 
Back
Top