We learned of single-PCB nVidia GeForce GTX295, we wondered about the purpose of the card. Most obvious one would be cost reduction, but there is a bigger white elephant in the room.
As we all know, GeForce and Quadro are mostly the same cards, but the difference is several million lines of code containing optimizations for professional applications. We all saw the adoption of GT200b chip through Quadro cards at first, and it is no wonder that nVidia plans to do a Quadro version of GeForce GTX 295. In fact, nVidia almost certainly has to do just that, since their direct competitor is going to launch dual-RV780 (could it be RV790?) based professional card in the next 10 days or so. With ATI most probably putting 4GB of GDDR5 memory on the card, nVidia had to react.
But the problem is that Quadro cards need to go through a grueling qualification process in order to become a certified part for supported professional applications. OEMs also have extremely high standards, and in this perspective, an enthusiast-designed GTX295 would have a real tough time.
The answer is very simple - build a single-PCB card, launch it as desktop part in May 2009 and take time to certify the design until first days in August and Siggraph 2009 conference in Louisiana. We spoke with a source positioned in one of Far Eastern companies, and we were explained that this would be "nothing out of the ordinary" when it comes to building professional graphics cards.
What makes things more interesting is the amount of RAM. nVidia gained quite a lot of experience with high-capacity GDDR3 chips, meaning that GeForce GTX295v2 would feature between 28 and 56 spaces for memory chips. If nVidia decides to populate all banks, current GDDR3 chips could increase the memory to 3.5GB, while going for all the 56 spaces could result in massive 7 GB of local video memory.
Usage of 64-bit operating system is a must for both cards, as FX5800 proved (card sees only 1.5-1.7GB on 32-bit OS, and all 4096MB on 64-bit ones). We will see if this rumor turns true, but 7GB of local video memory would enable 64x AA by using just one card. If two "FX4900X2" or "FX5900X2" would be paired in Quad-SLI setup, it could pave way for 128x AntiAliasing. But 14GB of GDDR3 memory in a single computer would definitely make sense for putting 48GB of memory in a single box, as our Nehalem-EP system proved.
If we talk about bandwidth, Quadro FX parts would most certainly have memory at 800 MHz DDR, nVidia's de facto clock of choice for Quadros (vast majority of Quadro line-up has exactly the same memory clock, Ed.), single Quadro FX 5900 X2 would come with 153.6 GB/s of local bandwidth. Two cards would have staggering 307.2 GB/s.
If you are a hard-core 3D app professional, you're free to start having wet-dreams about this one. Enthusiasts might ponder how come that their GTX295 is not passing certifications and forcing the company to develop a whole new design?
Only time will tell.