Folding@home, Not So Well...

By Matthew Dublin

At the recent SC09 meeting in Portland, OR, Stanford University's Imran Haque and Vijay Pande gave a presentation on the problems inherent in GPGPU computing due to a lack of error checking and correction in the memory systems of many graphics cards. They presented the results of a large-scale GPU error rate assessment test using MemtestG80, a software tool for assessing memory error rates on NVIDIA G80-architecture-based GPUs. While testing consumer-grade and dedicated-GPGPU hardware in a controlled environment resulted in no errors, results were dramatically — and perhaps alarmingly different — for consumer cards installed in real-world environments. After running MemtestG80 on roughly 20,000 hosts on the Folding@home network, they found that the majority of consumer-grade cards demonstrated what they called a "non-negligible, pattern-sensitive rate of memory soft errors." In other words, the data coming off of the Folding@home grid could possibly be in need of some serious quality control.

It's really good news!

It's really good news!