Two independent research teams from UNC Chapel Hill, Georgia Tech, Purdue, and other universities have demonstrated that Nvidia GPUs with GDDR6 memory are vulnerable to Rowhammer-style attacks that can give attackers full control of the host system. The findings, published April 2 and set for presentation at the IEEE Symposium on Security and Privacy in May, reveal attacks called GDDRHammer and GeForge that produce 64 times more bit flips than previously known GPU memory exploits.
The attacks work by rapidly hammering adjacent memory rows in GDDR6 chips to corrupt GPU page tables — the internal maps that control memory access. By manipulating the GPU memory allocator, attackers can forge page table entries to gain arbitrary read and write access to both GPU and host CPU memory, ultimately obtaining a root shell on Linux systems. Testing across 25 GDDR6 GPUs, including RTX 3060 and RTX 6000 models, showed hundreds to thousands of exploitable bit flips.
The vulnerability primarily threatens cloud computing environments where multiple users share GPU resources. Enabling IOMMU in the BIOS or activating Error Correcting Codes on the GPU can block the attack, though both options carry performance costs. GPUs with GDDR6X and GDDR7 memory are not affected.
