NVIDIA has alerted its users to enable System-level Error Correction Codes (ECC) to protect against GPUHammer, a variant of the RowHammer attack on its GPUs. This attack allows malicious users to corrupt data in GPU memory by inducing bit flips, significantly undermining the accuracy of AI models. RowHammer exploits the physical behavior of DRAM memory, contrasting other attacks like Spectre and Meltdown. Despite mitigations like target refresh rate, GPUHammer remains a threat, emphasizing the persistence of hardware security vulnerabilities.
NVIDIA is urging customers to enable System-level Error Correction Codes (ECC) as a defense against a variant of a RowHammer attack demonstrated against its graphics processing units (GPUs). This attack, dubbed GPUHammer, can cause malicious GPU users to tamper with other users' data by triggering bit flips in GPU memory, demonstrating a significant risk to user data integrity.
The most concerning consequence of this behavior, research indicates, is the degradation of an artificial intelligence (AI) model's accuracy from 80% to less than 1%. This drastic reduction in model performance highlights the critical impact that these memory vulnerabilities can have on AI applications.
RowHammer exploits the physical behavior of DRAM memory by causing bit flips in adjacent memory cells through repeated access. Unlike Spectre and Meltdown, which exploit speculative execution in CPUs, RowHammer poses risks unique to modern memory architectures.
GPUHammer is the latest variant of RowHammer, capable of inducing bit flips in NVIDIA GPUs despite existing mitigations like target refresh rate (TRR), showcasing the evolving nature and persistence of security vulnerabilities in hardware.
Collection
[
|
...
]