Source: The Next Web
University of Toronto researchers successfully demonstrated a Rowhammer attack against Nvidia GPUs even with error correction enabled—the mitigation Nvidia itself advised users to deploy. This exposes a gap between vendor guidance and actual vulnerability patching. Recommended defenses can be incomplete, forcing enterprises managing compute-intensive workloads (AI training, cloud processing) to assume their hardware may be compromised despite following official security protocols. GPU manufacturers may need deeper architectural changes rather than software workarounds to address memory-access vulnerabilities that have plagued semiconductors for over a decade.