New Rowhammer attacks give complete control of machines running Nvidia GPUs
15 hours ago
- #GPU security
- #Memory vulnerabilities
- #Rowhammer attacks
- High-performance GPUs costing $8,000+ are often shared in cloud environments, making them targets for attacks.
- Two new attacks exploit Rowhammer techniques on Nvidia GPUs, allowing malicious users to gain full root control of host machines by inducing bit flips in DRAM memory.
- Rowhammer attacks involve repeated access to memory, causing electrical disturbances that flip bits from 0 to 1 or vice versa, initially demonstrated on DDR3 DRAM in 2014-2015.
- Over the past decade, Rowhammer has evolved to target DDR3 with ECC, DDR4 with protections like TRR, use advanced techniques like RowPress, and even work over networks or root Android devices.
- Last year, Rowhammer was first shown to work on GDDR DRAM in Nvidia GPUs, achieving modest results with only eight bit flips, primarily degrading neural network output.
- The article notes it does not cover HBM memory, which uses extra stacks and parity for ECC, similar to having an extra DRAM chip, and HBM controllers manage complex capacities by disabling only faulty dies instead of entire stacks.