Researchers from the University of Toronto have uncovered a new attack called GPUThor that exploits Rowhammer to induce bit flips in the memory of certain NVIDIA GPUs, despite ECC error-correction protection being enabled. The researchers showed that the attack can cause a denial-of-service condition and, in a specific scenario, corrupt GPU memory page tables, allowing an unprivileged CUDA program to access arbitrary memory and open a root-privileged session on the host system, according to the research paper’s findings.
What distinguishes GPUThor?
Rowhammer attacks rely on repeatedly accessing memory rows to increase the likelihood of changing the state of bits in adjacent rows. The researchers modified the access pattern in GPUThor to make it irregular, thereby avoiding activation of the Target Row Refresh, or TRR, mechanism in GDDR6 memory. This modification was based on two undocumented behaviors in GPUs related to memory-request merging and the frequency of TRR activation.
The attack was tested on NVIDIA Ampere GPUs equipped with GDDR6 memory: the RTX A4000, RTX A4500, RTX A5000, and RTX A6000. According to the researchers, the new technique produced up to 6.6 times as many activations of attacked memory rows as previous approaches and recorded between 72,000 and 377,000 bit flips per gigabyte when ECC was disabled. These results represent an increase of between 4,548 and 23,597 times compared with the earlier GPUHammer attack, approaching the rates of powerful Rowhammer attacks on central processing units.
Why is ECC alone insufficient?
NVIDIA uses SECDED ECC to correct single-bit errors and detect two-bit errors within monitored memory blocks. However, with ECC enabled, the researchers recorded approximately 387 double-bit errors, which the mechanism can detect but not correct, along with two triple-bit errors that the study said ECC corrected incorrectly, resulting in data corruption.
Based on GPUThor’s rates, finding an exploitable bit flip could become possible in approximately 1.1 minutes, compared with approximately 21.9 hours with GPUHammer. This shift is the core significance of the research: existing protection does not eliminate the risk when errors move from the single-bit range to multi-bit patterns.
GPU disabling and privilege escalation
The researchers showed that GPUThor can place an ECC-equipped RTX A6000 into a denial-of-service state, causing the GPU to restart itself every two hours and terminate workloads. After the attack is repeated against the same GPU, it may eventually be classified as requiring replacement.
The more serious scenario involves corrupting GPU page tables. The researchers say this could give an unprivileged CUDA program the ability to access memory arbitrarily and open a root shell on the host system. They also noted that privilege escalation may remain possible on server-oriented Ampere GPUs, including the A100, because they rely on the SECDED level of ECC, even with restrictions that limit some denial-of-service effects.
What changes in practice?
The researchers notified NVIDIA of the findings on April 29, and on August 21 the company published guidance to mitigate the risk. NVIDIA recommends enabling SYS-ECC and IOMMU/DMA isolation, monitoring GPU error data, and restricting or controlling the execution of untrusted workloads. According to the company, susceptibility also varies depending on the type of DRAM, memory technology, platform design, in-memory protection mechanisms, and system settings.
No bit flips were observed in the same tests on GDDR6X or HBM2e modules, but the researchers warn that HBM3/e and GDDR7 modules with on-die ECC may also be vulnerable if multi-bit flips can be generated. The study recommends avoiding sharing a GPU between tenants whenever possible, monitoring ECC error counters, and restricting untrusted CUDA workloads. According to the researchers, complete protection remains dependent on developing stronger ECC for multi-bit errors and additional hardware defenses.