The implications are two-fold: the attack can cause a denial-of-service (DoS) condition by corrupting critical data, or it can be used for root-level privilege escalation. This undermines the long-held assumption that ECC memory is a robust defense against Rowhammer-style disturbances.

Key affected configurations include GPUs where ECC is enabled by default:

• NVIDIA H100, A100, and A800 data center GPUs
• NVIDIA RTX A6000 and RTX A5000 professional GPUs
• Consumer GeForce RTX 3090 and RTX 4090 (where ECC is explicitly enabled)

The research team demonstrated the attack without relying on any software vulnerabilities in the NVIDIA driver stack. Instead, they used a method called unprivileged GPU memory allocations, paired with massive memory traffic, to trigger the bit flips. They also confirmed that software mitigations like NVIDIA’s driver-level error reporting do not stop the attack, as the ECC correction itself is silently bypassed.

For administrators, the practical risks are real but require local access or a GPU-accelerated workload. Mitigation is not straightforward: disabling ECC is not a safe alternative, and the attack appears to be resistant to current driver patches. The researchers suggest monitoring for unusual error rates in GPU logs, though this may not catch the targeted bit flips.

Source: BleepingComputer

Is your organization running any of the listed NVIDIA data center GPUs, and if so, what are you doing to monitor for potential Rowhammer attempts on your GPU memory?