So I actually decided to read some of those slides this time instead of just looking at the pictures
here (Expreview). Is it me or does Cayman sound a lot like a AMD cpu with maybe some Bulldozer references thrown in?
Things like:
Asynchronous memory dispatch (like unganged?)
Dynamic clock control of "blocks" (similar to how cpu cores clock independently?)
Integrated power control processor (almost exact same name Intel uses to control Turbo)
Does the VLIW4 have two 64 bit Integers and two 32 bit Floating Point units?
Is is able to work on both at the same time (i.e. do both FP and Int calcs on same clock, ala Bulldozer)?
If this
is all true then Cayman looks more Bulldozer ready then Bulldozer (aka production ready). I can also see Cayman beating GTX 580 too, easily.
Also what's the significance of being able to fetch directly from the LDS? Are they possibly talking about the memory controller fetching directly from the LDS on the SMIDs? Also the slide (GPU compute) shows 8 "lines" connecting the first three steps before the shaders where as with Cypress it was only 1. Is there significance there or is it just more detail that was already there (like Barts showing the dual engines Cypress had even though the Cypress pics never showed it)?
I can't wait till 28nm! Sounds like it could be truly epic.
Even though the Cypress picture makes it look like it only has 1 dispatch processor it is actually like Barts in that there are two dispatch processors. Some of the Barts reviews covered this (I think maybe Tech Report & PCPer).