GPUThor Rowhammer Attack Bypasses ECC on NVIDIA RTX A6000

Academic researchers have disclosed a Rowhammer attack, dubbed GPUThor, that targets NVIDIA workstation GPUs equipped with GDDR6 memory. This novel attack successfully bypasses error correction codes (ECC), a primary mitigation strategy recommended by NVIDIA against GPU Rowhammer vulnerabilities. The exploit enables attackers to achieve denial-of-service (DoS) conditions and escalate privileges to gain root shell access on the affected host system. The research was conducted by scientists at the University of Toronto. They demonstrated the attack by hammering four DRAM chips on an NVIDIA RTX A6000 workstation GPU. Rowhammer is a class of hardware vulnerabilities that exploits the physical properties of DRAM cells. When memory rows are repeatedly accessed in rapid succession, charge leakage can occur, causing bit flips in adjacent rows. Historically, Rowhammer attacks have primarily targeted system RAM (DRAM). However, this new research extends the threat to the Graphics Double Data Rate (GDDR) memory found on GPUs. The researchers detailed their findings in a paper presented at the 33rd USENIX Security Symposium, held from August 11-15, 2024, in Philadelphia, Pennsylvania. The attack leverages the specific architecture and memory controller implementation of NVIDIA's workstation GPUs. By carefully orchestrating read and write operations to specific memory rows, the researchers were able to induce bit flips in critical areas of the GPU's memory. These bit flips, when strategically placed, can corrupt data or instructions that are essential for the GPU's operation and the host system's security. The ability to bypass ECC is particularly significant because ECC memory is designed to detect and correct single-bit errors, and in some cases, detect double-bit errors. This feature is intended to provide a higher level of reliability and security for critical systems, including professional workstations and servers. The GPUThor attack demonstrates that even with ECC enabled, sophisticated Rowhammer techniques can still be effective. The researchers' success in achieving root access means that an attacker could potentially take complete control of the compromised system, execute arbitrary code, steal sensitive data, or install persistent malware. The NVIDIA RTX A6000 is a high-end professional graphics card used in demanding applications such as artificial intelligence development, scientific simulation, and content creation. Its use of GDDR6 memory and ECC highlights its suitability for critical workloads where data integrity is paramount. The disclosure of GPUThor raises concerns about the security of professional computing environments that rely on these powerful GPUs. The University of Toronto researchers have provided detailed technical information about the attack, including the specific memory access patterns and timing required to trigger the bit flips. This information is crucial for NVIDIA and other GPU manufacturers to develop effective countermeasures. The implications of this research extend beyond the RTX A6000, as other NVIDIA GPUs utilizing GDDR6 memory with ECC may also be vulnerable. The academic community and cybersecurity professionals will be closely examining these findings to understand the full scope of the threat and to develop robust defenses against GPU-specific Rowhammer attacks.