For those of us concerned about packing maximum density into minimum slots, Galaxy is announcing a single slot Fermi card. Full article at AnandTech, at:
For people who aren’t limited by PCIe bandwidth, this could be a great way to get the best bang for the buck by minimizing the expenditure on non CUDA card parts. Additionally, it could allow a standard 7 slot ATX or ATX mid case to have 4 cards in it, either by packing in 3 others and this one, and/or allowing more space between cards for better airflow.
Naturally, one is right to be concerned about the temperatures such a card would reach. Let’s hope this issue isn’t a dealbreaker.
However, you slice it, this opens up several interesting possibilities.
Assuming these don’t completely melt in high-density packing arrangements, this means you could use one of the 8-slot multi-NF200 motherboards out there to roll your own quad “GTX 495”-like system without having to wait for NVIDIA to release that card.
I am pretty sure we won’t get to see one of those this time around. The current Gforce Fermis run in the 230-250 watt range. There is a limit of 300W per device in the current PCI-e 2.0 specification, so unless they find a bunch of TDP savings in a respin, or unless they use a seriously underclocked part, probably with fewer than the current minimum 448 cores, or unless they raid the top bin, or some combination of all three, I don’t see how a PCI-e 2.0 compliant “GTX495” is going to happen.
Speculation at this time is that if we see a fermi dual GPU solution it will probably be build off the GTX 460 design and be underclocked to squeeze by power issues and heat buildup. I myself am trying to figure out what the best number cruncher will be from this series for the next phase of my folding farm.
I am currently trying to find someone to help me figure out whats the best linux distro for running the SMP clients and GPU clients with a minimum of headache. If anyone can point me at someone that may be able to help baby walk me through this it would be greatly appreciated.
Yall may also be interested in some of the case designs I am currently working on that provide better cooling than open air U4 setup for mutiGPU farms. many of the things that will work for my projects will also be useful to yall as well I believe.
The GTX260 version of the Galaxy Razor was on the market for maybe one week last winter before disappearing. Since you could get the same GPU density from a dual-GPU Fermi card, ala the GTX295 (edit: unless a “GTX495” is unfeasible as avidday suggests), the real value for a single slot Fermi would be as a retrofit for 1U cluster compute nodes. You would need blowers positioned properly, but this should be doable.
Well, during the GT200 product cycle, NVIDIA first released the GTX 280 and GTX 260 on a 65 nm process, running at 236 and 202W, respectively. Then they switched to the 55 nm process and 6 months later and we got the GTX 285 and 295. I expect that this cycle might run similarly. Maybe next year (probably not 6 months given how tough 40 nm was) we’ll see 32 nm parts and a dual GPU solution.
We won’t see 32 nm GPUs because TSMC cancelled their 32nm bulk node more than 6 months ago, and it looked like there won’t be any 28nm bulk production for at least 15 months, possibly longer. Anything that NVIDIA and ATI bring out in the next year will also be on 40nm. I very much doubt any Fermi based part will ever be fabbed on anything other than the current TSMC 40nm process, and therein lies NVIDA’s problem.
I agree, and thats whats going to limit our options. It will be quite awhile before we see a refresh of the 400 series because of issues they have experienced in the die shrink process. Best we can hope for the moment is that the manufacturers can come up with better cooling solution for the GTX 480.