Search
"GPUHammer": Silently Corrupting AI Models on NVIDIA GDDR6 Graphics Cards
2 day ago /
Read about 0 minute
Author：小编
A team of researchers from the University of Toronto has recently unveiled a novel attack method dubbed "GPUHammer," capable of covertly disrupting AI models running on NVIDIA GDDR6 graphics cards. This attack causes the accuracy of the models to plummet from 80% to a mere 0.1%. Leveraging the Rowhammer principle, "GPUHammer" repeatedly reads and writes memory cells, thereby inducing electrical interference and altering data in adjacent memory rows. In response, NVIDIA has provided guidelines advocating the activation of ECC (Error-Correcting Code) functionality to mitigate these risks. However, enabling ECC may come at the cost of a 10% performance decline and a 6%-6.5% reduction in video memory capacity.
-
C114 Communication Network -
Communication Home
7 X 24 Track global technological trends
