RV515 got a Ring-Bus(?)

no-X

Veteran
Trying to find differences between mobility Radeons, I noticed that mobility X1300 implements ring-bus. X1300 is 100 mil. trns. and X1300 mobility (only) 105 mil. trns.

Mobility™ Radeon® X1300 Product Features

* 105 million transistors using 90nm fabrication process
* Four pixel shader processors
* Two vertex shader processors
* 128-bit 4-channel DDR1/DDR2/GDDR3 memory interface
* Native PCI Express x16 bus interface
* PowerPlay 6.0 power management technology
* Avivo Video and Display architecture

Ring Bus Memory Controller

* 256-bit internal ring bus for memory reads
* Programmable intelligent arbitration logic
* Fully associative texture, color, and Z/stencil cache designs
* Hierarchical Z-buffer with Early Z test
* Lossless Z Compression (up to 48:1)
* Fast Z-Buffer Clear
* Z/stencil cache optimized for real-time shadow rendering

http://ati.amd.com/products/mobilityradeonx1300/specs.htm

all mobility X13x0, X14x0 and even new X2300 have ring-bus and consist of 105MT. So, we know, that RB in R520 (and probably in all x-1-1-1 chips) is about 10% of the die. Mobility X1300 is 5% bigger (in transistor count), than RV515. I can say that (at least 256bit) RB is 5% more expensive, than traditional memory controller. Using extrapolation, R520 w/o ring bus would be about 15MT smaller.

(a bit of OT - 1024bit ring-bus probably consumes 30-45MT(?) more than traditional memory controller. And this number is close to the difference between R600 and G80 - other way - the rest of both cores consist of very similar number of transistors and it will be easier to compare effectivity of these two architectures.)

Anyway, the question (possibly answered) is why former RV515 didn't have RB...

Jawed said:
Seems like an interesting development, implying that RV515 didn't have a crossbar because of die area, but because of the development timeline.

...and why ATi implemented it additionally? Could be this the reason?

Dave Baumann said:
...numbers from states Virtual Memory for R520 (and "NO VM" for RV515)
http://www.beyond3d.com/forum/showpost.php?p=703708&postcount=2

It seems to me, that ATi wants virtual memory support for all new products because of Vista...
 
Last edited by a moderator:
I have to say I'm mildly surprised that the ring-bus is appearing in these "smallest" dies. But if it enables virtual memory support then the cost of the ring-bus over a crossbar is even lower than I presumed.

So it seems likely that even the smallest R6xx GPUs (around 80mm2 if I remember correctly, smaller than any RV5xx) will have a ring bus.

That old thread is quite interesting. Pity it just stopped like that. I do wonder whether Vista can "brute force" virtual memory onto GPUs that nominally "don't support it", such as the old discrete X1300. This would require driver support and presumably has limited performance/capabilities.

rwolf made a very interesting point:

http://www.beyond3d.com/forum/showpost.php?p=704460&postcount=29

What will be interesting is how VM will be handled with SLI and Xfire.

and I'm afraid I can't help thinking that maybe this is at the heart of NVidia's troubles getting SLI working under Vista.

Anyway, here's two ATI patent applications relating to virtual memory:

Methods and apparatus for updating a memory address remapping table

Method and apparatus for fragment processing in a virtual memory system

The latter appears to answer the question of memory fragmentation: it doesn't matter.

Jawed
 
I'll probably have to ping the notebook marketing guys again.

Along the lines of "Quit that, already!" ? Taking a generic X1k family characteristics document a little too far in reprinting them for marketing purposes?
 
Back
Top