phongn wrote:Well, the CPUs do dynamically overclock so it might make some sense. ECC should protect against some errors during overclocking, too?
ECC won't help prevent errors in CPU overclocking. Theoretically it might with memory overclocking, but I doubt it helps much otherwise all the gold-plated super-enthusiast memory (and people going for overclocking records) would use ECC. I'm not a hardware engineer, but my guess is that timing skew errors are the primary problem with memory overclocking and ECC doesn't really help with that (because it tends to cause command/address corruption and multi-bit errors). In any case all the available ECC DDR3 memory is considerably slower than good non-ECC memory, so I won't be using it for this machine.
Couldn't you do a 2TB RAID-1
Simple RAID 1 is almost never done with more than two disks, because the capacity of the array is only the capacity of the smallest disk in the array. Given four disks, the usual solution is to aggregate them into two large volumes with RAID 0 and then RAID 1 mirror that (or vice versa, it doesn't make a lot of difference).
(or even a RAID 5) array?
I'd need a controller card or software RAID for that, the motherboard doesn't support it. RAID 5 is neat, I'm running it on my file server, but I'd rather go for the simpler solution here.
It just seems weird that you're going for RAID 10 with slow drives
It's a simple way to get redundant storage. It's a little less cost-efficient than software RAID 5 of 3 drives but really the cost is trivial compared to the rest of the machine. Now that I think about it, perhaps I should ignore RAID for the HDs and leave them as separate volumes, then set the backup script to mirror the SSD array to different drives in sequence. Or I could use 3.5" drives in external cases (the motherboard has a couple of ESATA ports), but that would be kinda messy. To be honest this is the least important bit of the machine.
What about using AMD's FireStream cards?
Not quite as hideously overpriced as the Nvidia Tesla cards, but still, lots more money for only one real advantage; 4 gigabytes of memory per GPU instead of 2. Which would be nice, but I'd rather have GPUs I can overclock by 20%. Remember that unlike Nvidia, AMD don't cripple down the dual precision performance of their consumer cards by 75%. Also I don't think standard waterblocks would fit FireStream cards.
They've got some better features for compute that were left out in their consumer cousins.
Well, currently I am running the application of interest on a 1 GB 5850 with no problems (albeit on relatively small document sets). How do you think a FireStream (or Tesla for that matter) would help?