Quote:
Originally Posted by Farhan
Hehe, i was allowed 4 metal layers.
|
Pfft, you young whipper snappers and your fancy tools. We had to do with a single metal layer! Let's just say it made place and route very interesting. And by interesting, I mean unnecessarily tedious.
Quote:
Originally Posted by silent_guy
Switching from defense to offense: other than some non-system related implementation details, is there a single advantage of a ring over a crossbar?
|
Considering that there are just up to 4 ring stops in R5xx and R6xx, wouldn't a point-to-point bus be even better? In terms of wiring/power, it's not that much worse than a ring, and it does have the advantage of drastically reducing the scheduling complexity, latency and additional buffering.
Of course, that doesn't scale well above 4 points, but then again, we haven't seen the "ring bus" do better yet.
Quote:
Originally Posted by _xxx_
But why the 512-bit bus then?
|
Because efficient non-power-of-2 memory addressing is hard. So if you need more than 256 bits, your next bus width becomes 512-bits. All IMHO. It would be interesting to see if anyone can write a test to max out the bandwidth on R600, given the rumours of 16 ROPs RBEs and 16 Textures.
Quote:
Originally Posted by Jawed
IBM went with a ring bus for Cell, hugely motivated by simplicity of implementation.
|
Cell also has 9 "stops" on their ring. That's somewhat more than 4. |
