With LCDs becoming much more affordable I just hope Nvidia Fermi comes with Mult-monitor support.
Then reviewers can do some memory bandwidth/VRAM comparisons at higher resolutions than 2560x1600.
Multi-monitor support existed for NV a long long time ago.
My 7800GT on XP could do 2560x1024 on two 17" monitors, any 3D app would work, Windows saw it as a single monitor of 2560x1024. And that was over 4 years ago.
Whether they still have it, I don't personally know, the important question would be how many outputs does an NV card have, and if you can span multiple monitors, can you use the outputs from multiple graphics cards in SLI (or even not in SLI, such as using a primary dual output card with a secondary PhysX card and spanning 3 monitors across the two cards).
Such things are possible, in theory and practice (or have been), since NV have managed dual outputs as a single screen in the past, and Software Triplehead (
http://www.kegetys.net/SoftTH/) has managed to give 3D graphics spanned over multiple displays on different cards.
Whether NV will actually implement such features is another matter entirely.
Unfortunately I did go off topic once again in this thread, so I will have to add something to get back on topic.
Back on to ECC:
Not much. The ECC support on the Telsa cards is a bit of a hack. While the Telsa card will have 6 GB of memory on it, only about 5.25 GB will be usable with the rest being reserved for its ECC implementation. Additionally the ECC algorithm utilizes bandwidth on the main memory bus and require their own load/store operations which will impact overall performance.
http://forums.vr-zone.com/news-arou...st-gpu-dual-gpu-fermi-coming.html#post7456148
Not sure 100% where that is from, but I was sure I had read something similar somewhere (not on a forum), and that seems to explain the ECC handling quite nicely.
Back to the question at hand, I just thought of something, although in some ways it's strange, in others it makes perfect sense.
NV has Fermi.
It's 3bn transistors on a 40nm process.
The HPC version will have 448 pipes, the desktop we assume is still on for 512.
Even at 448 the Tesla card will almost certainly blow anything else totally out of the water.
What's to stop NV using defective harvested die for the Tesla cards, and using fully functional ones in desktop graphics cards?
Sure, it's a bit different to normal (cut down/defective die being used as lower end cards), but if that many aren't making the 512 cut, then maybe it's a sensible move to harvest the maximum number of reasonable dies by setting the Tesla high end bar at 448, and having the desktop as 512 and 448, meaning that you're not so pressed for Tesla GPUs?
Does that make sense?