entee writes: Researchers at Stanford tested over 20,000 NVIDIA GPUs around the world using the Folding@home distributed computing platform, and found that a substantial number of NVIDIA GPUs exhibit "soft" memory error. Even after applying a crude correction for computing errors, as many as two-thirds of graphics cards were found to "forget". The paper claims this is a problem in all existing NVIDIA GPUs: "G80 and GT200 are both vulnerable to a nonzero per-transaction error rate in memory accesses", and that it is likely true of boards from all GPU manufacturers. This study comes on the heels of Google's recently reported "hard" memory error rates in its server farms. Given the rapidly swelling number of GPUs user for scientific computing applications, perhaps this result explains why NVIDIA plans to introduce ECC into its next generation hardware.
Dan Cabrera writes: "So I just took delivery of a large package (pickup truck bed sized box) labeled 'server' — turns out it's the one I've been waiting for a while from a client. We had some miscommunication and I assumed this was a 2-4, maybe 5TB system for use as a production content shared drive, but it turned into a real monster, SuperMicro SC-836 with HighPoint 2240 Controller and *16*x*1TB* Hitachi drives in a RAID5 array. It's got WinXP64 loaded, but there must be a better solution after reading of ZFS and related technology now available as open source, no?
Looking forward to your comments and suggestions! I'll repost with some benchmarks as this puppy grows up (into the wee hours of the morning I'll work:)
PS: Just need a large drive to save rendered projects to and backup music/other projects (This is for a world known DJ, so the each tour/show can involve a LOT of media!)...and, how the heck am I gonna do off-site backup for this in event of disaster? Ay-yay-yi!"