Dual ATI Card Finally!

Not sure what you mean about physical requirements.

2 x 96 shader units is pretty impressive, that's 50% more per card than a single R600 which has 64 (4+1) shaders. If it's not a typo that gonna be a ridiculously power consumptive card wihout fixing the 80nmHS issue IMO.

Also WTF its the card if it' 96 shaders each, that alone would be more than the HD2900XT, and likely an 850+million transistor part.

If it's a typo and a combined 96 shaders (48 each) then that very well could be the HD2900Pro we've been hoping for, but unfortunately not on 65nm, which would be a dissapointment.

If it's 96 in dividual shader element units / stream procs, then that's worse than a single HD2600, and would struggle IMO to do much of anything.

Seriously this information from the InQ leaves me with more questions than it answers, and I don't trust the impression I get from what is written. Something seems very wrong on it.

Awww I didnt see that either


Topic is worthless then I guess

Yep already aknoledged that.

The dual chip 2600xt has been known for quite a while, unfortunatly it only has 2x 128bit memory controller, which is, quite crap. Will be intersting to see if it competes with the upcoming 2900pro 65nm though.


And according to a few sites ATI could POSSIBLY make a dual 2900pro chip. Now that would be something to look forward too for once!

There was also talk of a dual RV670 (HD2900Pro) card which is suposed to be 256bit, to be AMD's combatant against the high end, but if that's not coming out until the Fall, I don't hold out on that being much of anything. Although it might give them some practive on working with a more modular R700.

However I wouldn't hold my breath for any of them, and I think AMD need to test their feet with making the HD2600 Gemini work well before relying on such a solution as a high-end contender. As we saw with the GX2, it's unlikely to sell enough for it to be a winner in the segment unless it's nigh on flawless.

That would be worth it, as would a dual 8600 from Nvidia. Sadly, after the latest Anandtech DX10 benchmark, I have little faith in the "midrange" cards really being midrange in DX10.

Seems to me the true midrange per performance are the 8800GTS, X2900XT and 8800GTX. No lesser card need apply; so why not "gemini" them to fill in that missing market slot?

How about a Dual 8850GX2 or 8900GX2

Oh and while there at it. They do make boards with 3ea PCIE slots how about 6ea cores running SLI

From what I've read the HD2600 Gemini will use DDR2 ram just plain DDR2 ram then again hopefully there might be variants of it eg DDR2 version for $189 GDDR3/4 for $250 yeh I know it probably wont be that cheap but thats all the info we have on the price. Also the HD2600 Gemini will require a Crossfire compatible mobo thats a bit of a downer.

I think the 3rd PCIe slot in Crossfire is for physics, not that I think it's been implemented yet. Something like two X2900XT's for Crossfire and an X2600 for physics would work.

Nvidia's pushing a 4 PCIe solution, but I'd hate to see the heat issues with that. What does it need, a regular 1000 watt PSU plus a 1000 watt GPU PSU?

That's why a dual GPU on one card is more tempting than regular Crossfire or SLI.

Just remember there are already mobos that support 4 PCIe16X PEG slots (although most at 8X electrical), so you could even have 8 cores depending on what the mobo requirements are.

And to Hatman, I don't know if they need Xfire MoBos or not, but they (ATi & nV) have required them on other implementations that really didn't 'need' it.

IMO the major constraint will be memory bandwidth for 4 card / 8 vpus.

If the ROPs from each chip have to write to each memory bank a copy of what each one has, I see a huge memory I/O problem brewing.

It's old news announced at Computex 2006, so I don't know what the current status of the implementation is. I think AMD and ATI had enough on their plates due to the merger and the X2900 series, so I don't know if any boards have actually arrived to market yet. I don't know if Havok physics on ATI GPU's is implemented in any games at this point.

http://www.trustedreviews.com/grap [...] Physics/p1

It's like Nvidia's GPU physics and new announcement of hybrid SLI, something to look forward to if it and when it pans out.

My MSI K9 board only supports x8 in the single PCIe, and some of the older ATI boards do to. Does that really affect performance of, say, an 8800GTX? I've noticed in Crossfire and SLI reviews that many boards only support x8 if two cards are used, but x16 if one is used. What's up with that?

Yes, SLI is 2 cards in x8 mode, but does one card in x8 mode on a single PCIe motherboard lose performance? Does x8 have enough of a data transfer rate for today's games?

So, far, I haven't noticed any issues with the 7600GS, but what about an 8800GTS or "9800GTS"? When will games require x16 PCIe?

The first link is useful, 2nd looks to be a load of rubbish. 230euros? No way. Thats £170/$340. The first link suggests about $189 to $249 which is a LOT more reasonable. It also cnfirs the source of 2 - 3 other sites taht say itll only be $30 - $60 more than a single 2600xt.

It's kinda silly you need a xfire board to use it... might aswell buy 2 of them if you're going to get it at all then.

At around $440 it would price about the same as 2900xt. Would be interesting to see if 4x 2600xt's beat a single 2900xt lol. If it comes in GDDR4 version it should easily beat the 8600gts.

No games on it yet, the latest thing was the smoke and liquid demos from ATi.
Havok showed a UBisoft game in March in SF, but no mention of what game it was or how impressive/unimpressive it was. IMO, no 'new' about it means it was a mediocre tech demo at best, showing they are working on it, but nothing impressive. There was talk of the Fable sequel getting HavokFX, but I don't think so, I think people are readng in too much to what's been said about it using Havok physics (which is their entire engine both CPU and VPU based). It's like the early hype about Oblivion which turned out to be untrue.

My MSI K9 board only supports x8 in the single PCIe, and some of the older ATI boards do to. Does that really affect performance of, say, an 8800GTX? I've noticed in Crossfire and SLI reviews that many boards only support x8 if two cards are used, but x16 if one is used. What's up with that?

Well remember that the chipset only spports so many PCIe lanes, so older chipset supported 20-24 lanes, there's no way for them to put 2 PEG slots there that are electrically 16X each, they need to be 8X, or if only one slot is used they could switch to 16X for the one slot (some even allow 16X + 1-4X allowing for multi-monitor [which usually disabled or at least didn't use the second card for gaming] allowng the second usually weaker card to still function with more than enough bandwidth for 2D and mild 3D).
Then later implementations got 30+ lanes, so you could easily use 24 for graphics, and then 4 for a 1-4X slot or multiple 1-2X slots, and then one to audio, one to ethernet, etc. and then you had 16X + 8X configurations, or like the some you got combinations of 16x, 8x and 4X, with some even being 8x+8X+4X+4X.

Then came nV's dual chips solution which boosted the lanes to about 42-44 between the 2 chips, and thus you had 16X+16X as a possibility. Then ATi increaed their support to the 40range giving them 16x+16X potential. But on the 3 slot board they were stil stuff at 16+16+8 because they still didn't have enough for more. And the quad slots were still combos fo 16X and 8X.

Now the newest boards from intel and AMD (nV soon to come) are able to support 3 16X lanes slots, or 16+8+16+8 depending on config.

Future board will liley support 64+ lanes allowing for 4 x 16X, but I thinnk that's alot of layers, and a heck of alot of communication going on.

Also rememebr that PCIe 2.0 will bring a speed boost and thus the wires for 16X will carry the equivalent info of 32X, so you'll still need the support in the chip, but future baord could basically have less lanes and support the same throughput.

IMO 8X is the bare minimum to achieve the maximum from the cards, but you need more if you add to the workload like doing CF/SLi cross-communication through the PCIe lanes instead of a dedcated bridge.

That being said though, and in response to your later question, it's likely that 2 cards in 4X slots runnning in CF/SLi will still outperform 1 running at 8-16X in apps that see improvement by CF/SLi, where the throughput isn't as limiting as the graphics power limit.

No it means 2slot cooling solution. Like many other cards that the HSF assembly is big enought that if there is another slot (like PCI or PCIe) right beside it, you won't be able to use it. It doesn't mean it requires another PEG slot which you can't use.

Most boards that anticipate CF/SLi have the slots arranged to fit the huge HSFs of cards like the GF6800U/X850XT to the GF8800 and HD2900 series.

Look at the original Quad Gigabyte and notice the staggering of PEG slot, then other slot, then PEG slot, etc. (second page has a better to-down view);
http://www.tomshardware.com/2005/1 [...] therboard/

Now to contrast that they also have a Quad which include single slot ones as well, if they think people don't have room, and will fill them with single slot cards;
http://www.bit-tech.net/news/2007/ [...] rossfire/1

So you have both options, but the one there talks about 2 slot cooling, not lane requirements.