NVIDIA Issues Security Notice on Potential Rowhammer Vulnerability in GPUs

In a recent security update, NVIDIA has addressed concerns raised by researchers about the potential for Rowhammer attacks targeting its GPUs (graphical processing units). The Rowhammer vulnerability has been known for several years, originally identified in CPU memory, but new research indicates that modern graphics cards, such as the NVIDIA A6000 using GDDR6 memory, may also be susceptible under certain conditions.

What is Rowhammer?

Rowhammer is a type of security vulnerability affecting dynamic random-access memory (DRAM). It exploits the physical properties of DRAM chips, where repeatedly accessing (or “hammering”) a row of memory can cause bit flips in adjacent rows. These unintended bit flips could, in theory, be manipulated to alter data or compromise system security.

Researchers from the University of Toronto demonstrated a successful Rowhammer attack on an NVIDIA A6000 GPU when System-Level Error-Correcting Code (ECC) was disabled. ECC is a feature that detects and corrects certain types of memory errors, serving as a primary defense against such exploits. The study highlighted that enabling ECC effectively mitigates the risk, underscoring the importance of proper system configurations.

NVIDIA in response reaffirmed that most of its current GPU and System-on-Chip (SoC) products incorporate robust memory controllers designed to meet prevailing industry standards. They emphasize that the actual risk of a successful Rowhammer attack depends on multiple factors including the specific hardware, system settings, and whether ECC features are active.

Mitigating & Best Practices

To combat potential vulnerabilities, NVIDIA recommends the following measures:

• Enable System-Level ECC – This is especially critical for enterprise and data center environments. Many NVIDIA GPUs, such as the Blackwell, Hopper, and certain Ampere series, come with ECC enabled by default, but users should verify this setting.
• Use Supported Hardware – Modern GPUs, especially those with DDR4, LPDDR5, HBM3, and GDDR7 memory, incorporate On-Die ECC (OD-ECC), offering an additional layer of protection against bit flips caused by Rowhammer.
• System Configuration Checks – Both out-of-band and in-band methods are available for verifying and enabling ECC on supported systems. For example, system administrators can use NVIDIA’s command-line tools or system management interfaces to ensure ECC is active.

While consumer-grade GPUs are generally less exposed to such risks, enterprise and data center users should prioritize security configurations:

• Confirm that ECC is enabled on all critical hardware.
• Keep firmware and driver software up to date.
• Consider segmenting GPU workloads to prevent multi-tenant access, which can increase attack vectors.

NVIDIA continues to monitor emerging research and emphasizes that proper system configuration remains key to safeguarding against these types of vulnerabilities.

As DRAM technology evolves, with newer standards like GDDR7 and HBM3, features like OD-ECC are becoming more prevalent, providing stronger defenses. And while the recent research highlights a potential vulnerability, it also underscores the importance of proper system setup. Enabling ECC and being informed on hardware security features are vital steps for organizations relying on GPUs including NVIDIA’s to maintain data integrity and security.

For more details or support, visit their advisory notice here.