10 comments

[ 4.7 ms ] story [ 33.4 ms ] thread
I've been wanting to learn the internal architecture of GPUs for a long time, so this will certainly be very useful. Thank you.
One set you might find interesting, with likely expired patents, are the once-revolutionary SGI systems. Their O2 workstations and InfiniteReality systems specifically. Systems like O2 and Octane were used for movies such as Fight Club plus visualization applications (and Quake). The InfiniteReality technology was used in the Final Fantasy movie albeit with a lot of it. These might be copied in an open effort.

Note: To be clear, these aren't GPGPU systems but straight GPU's. I think MIMD architectures for general-purpose have proven better than GPU's. Just less investment and good one's mostly get acquired. FPGA's can be better, too, with good HLS tools and S-ASIC/ASIC conversion for better performance later.

http://webstaff.itn.liu.se/~matco/TNM053/Papers/p45-kilgard....

https://en.wikipedia.org/wiki/InfiniteReality

Far as internal architecture of regular GPU's, just search for it in Google with terms like that to find it. There should be descriptions of their architectures and plenty that came before them.

Since this is based on AMD's GCN architecture, aren't there any patent issues?
the ISA isn't open?
What do you mean by open and what does that have to do with patents?
Now, if you add a realistic memory hierarchy model, it can be used for a far deeper performance analysis than the official OpenCL tools are providing.
UWM = University of Wisconsin Milwaukee.

The University of Wisconsin Madison is referred to as UW, "Wisconsin", or The University of Wisconsin -- but not UWM.

Isn't University of Wisconsin Milwaukee the folks that have the really good cryptography courses?
related: AMD published source code of HSA Runtime library

https://news.ycombinator.com/item?id=8605764

________________________________

and yes, it's Univ Wisconsin-Milwaukee, where i used to ride my bike to take my first programming classes.

It should be noted that this is the design for a single "Compute Unit" of a GPU, not an entire GPU. A GPU consists of an array of many of these connected up in a complex fabric, along with a lot of other components. The linked design includes none of these.