GPUThor is an advanced Rowhammer attack developed by University of Toronto researchers that targets Nvidia Ampere GPUs (models A4000, A4500, A5000, A6000) equipped with GDDR6 memory. It improves on previous GPU Rowhammer attacks by exploiting the TRR mechanism's refresh cycle limitation, triggering bit flips at a rate 7,000 to 23,000 times higher than prior methods. This results in 72,000 to 377,000 bit flips per gigabyte, enabling double and triple-bit errors that ECC memory cannot fully correct. The attack can cause denial of service by forcing GPU reboots and hardware faults, although arbitrary code execution remains unproven. The research highlights the ongoing risk of Rowhammer attacks on GPU memory, especially in cloud environments where GPUs are shared resources.

The attack can cause significant data corruption in GPU video memory, resulting in denial of service conditions such as GPU reboots and hardware faults that require replacement. ECC memory protection is insufficient against double and triple-bit errors induced by GPUThor. Although arbitrary code execution has not been demonstrated, the increased effectiveness of the attack suggests a theoretical risk to cloud infrastructure and industrial GPU deployments. No confirmed exploits in the wild are reported, and the attack remains largely theoretical and experimental at this stage.

No official patch or fix is currently available for GPUThor. Standard ECC memory protection is insufficient against the double and triple-bit errors caused by this attack. Cloud providers and GPU users should monitor vendor advisories for updates. Given the attack's theoretical nature and lack of demonstrated arbitrary code execution, immediate urgent remediation is not required, but awareness and cautious use of shared GPU resources are advised. Patch status is not yet confirmed — check the vendor advisory for current remediation guidance.