]
the ATI xpress3200 chipset boards has dual x16 port, along with a few adtional PCI-E connectors for gigabit, and the southbridge
yes, that's still ONLY 32 lanes per controller, 32 from the north bridge and 4 or 8 from the south bridge :eek: isn't that amazing
VIA Technologies, nVidia, and Intel, ALL have competitive (and imo better) chipsets that accomplish the same thing
http://www.via.com.tw/en/images/products/chipsets/k8-series/k8t900_blkdiagram.jpgObserve, 32 lanes are provided by the northbridge maximum (in this case it can provide a total of around 19) and an additional 2 or 3 provided by the southbridge
a GPU, to stay competitive with at least AGP 4x (which is the rough absolute MINIMUM bandwidth you can have between a GPU and system and maintain any sort of respectable performance, and AGP 4x's ability to function as such is slowly detiorating) needs roughly 4 PCI Express lanes, 16 provide it with around 4GB/s of full duplex bandwidth, AGP 8x, by comparison, is 2.133GB/s of half duplex bandwidth
so PCI Express x8 provides roughly 2GB/s of full duplex bandwidth, and PCI Express x4 provides roughly 1GB/s of full duplex bandwidth (which does have it's advantages over an interface like AGP 4x)
So, in a system like the Quadro Plex Model II, which features 4 GPU's, each GPU only gets x4 bandwidth, 8 times over if you're running 8-way SLI, now, you might be thinking, ZOMG, SO LITTLE, then consider that nVidia's created a massively parallel rendering scheme here, and you probably won't even be utilizing 1GB/s to each GPU (considering a conventional GPU can render full frames with 1GB/s and usually not have a ton of problems)
now after this little lesson on PCI Express and north bridges has been given, my original point was merely that each one of these things requires a very large piece of bandwidth to operate correctly, and given that nVidia is advertising these for rack configuration w/TWO for a single rack (width wise) that means each pair needs (basically) 32 lanes (although i'm sure if you played with drivers/intelligence you could get two to work on dual x8's, or if, for example, your motherboard only provided a total of 16 lanes, such as the original NF4 SLI, or the Intel 925x) SO! my original point is furthered to a question asking, how the hell would you build an IG out of these, given that PCI Express is A) a local interconnect and B) limited to 32 lanes per controller, and 32 lane controllers are rather expensive, a layout similar to Gigabyte's Quad Royal could be pulled off, but that could still only provide maybe at most 64 lanes
Quad SLI was speculating something like PCIe x40 on the original presentation, however it's being released to consumers to run on dual x16's, which is still inside that 32 lane configuration...in other words, to design a system capable of running more than a pair of these in tandem (lets just ignore the whole driver issue, the whole heat issue, the whole power issue, and for the sake of time and our sanity, the whole issue of physical distance's effect on a local interconnect's latency in this case...ummkay) would be a total pain, as you'd need at least 64 available lanes (so copying Gigabyte's Quad Royal configuration, you link a pair of NF4 or NF5 northbridge's together, Gigabyte linked the Intel Ed. to the AMD edition, as the AMD edition has no memory controller, so it can function as an SB, it provided a full 32 lanes in a 4 x8 config or 2 x16 config (they were gonna release quad SLI drivers for their 3D1 cards with it, but I guess that never happened, sadly, I wanted to see 4x6800GT, the performance would still be formidable))
ANYWAYS
considering the above, and considering the total worthlessness of this for an IG (it isn't designed like a true IG, like something from E&S or the old Quantum3D machines from the late 1990's that ran on VSA-100's) given the PCI Express thing, WHY WOULD YOU MAKE IT RACK MOUNTABLE, as you'd most likely never wanna stick something like this in a data center, and it makes no sense for a graphics workstation to sit in a rack (unless it's a small rack, similar to a DJ station, however a 43U rack would be totally overkill for a machine that shouldn't be much larger than a desktop)
The only advantage to the "rack" config I can see is if your workstation is a desktop and is similar in durability to the SGI Indigo2 (or similar) series, so you could stick your little 8 GPU IG farm on top of it, nonetheless, this technology seems almost entirely over the top, then again, it's the world of progl cards, what do you expect
