Skip to content
Business Efficiency Talking point

Black Hat researchers say Rowhammer worsens as chips get denser; NVIDIA GPUs are exposed

Researchers say denser memory is more exposed, and ECC, NVIDIA's recommended fix, costs speed and capacity.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Black Hat researchers say Rowhammer worsens as chips get denser; NVIDIA GPUs are exposed

Graphic: WebPulse, drawn from the sourced facts in this story.

In brief
  • Black Hat researchers said Rowhammer, a memory flaw about ten years old, grows worse as chips get denser, and newer DDR5 memory is still affected.
  • NVIDIA's recommended fix, ECC, costs memory speed and capacity. The speakers said treating the GPU as untrusted stops their CPU-side attacks.

Researchers at Black Hat USA 2026 argued that Rowhammer, a memory flaw about ten years old, is getting worse as chips get denser. Their briefing on GPUBreach was published on November 19, 2026. It describes an attack on NVIDIA graphics chips. The excerpts do not name the speakers.

What was said

Rowhammer is a physical effect. Reading memory rows very fast causes electrical disturbance. That flips bits in nearby rows between 0 and 1. A speaker asked whether a flaw this old should be fixed by now. The answer was no. The problem had grown, and even the newest DDR5 memory remained susceptible. In their words: "And as our chips evolve more, they get more vulnerable to Rollhammer after each generation as they get denser."

Earlier work flipped bits in NVIDIA GPU memory and degraded machine learning accuracy. GPUBreach goes further. The show notes call it the first targeted Rowhammer attack on NVIDIA GPUs that allows privilege escalation, meaning higher access than intended, and cross-process memory access.

To get from an ordinary CUDA program to root, the researchers first studied how the GPU lays out its page tables. A speaker called a page table the address book of a GPU. The defenses they listed hint at the route: keep page tables apart from ordinary data, so hammered rows cannot reach them. They also warned: "CPU driver is not safe from a malicious GPU."

Why it matters

Our reading: this is a recurring bill that comes from memory physics, and software cannot fully erase it. A speaker recalled an NVIDIA bulletin that lists more than 20 of its product lines as exposed. NVIDIA recommends ECC, which detects and corrects some bit errors. The speakers said not every GPU has it, it hurts memory performance and capacity, and it has long been shown not to be foolproof. They gave no size for the cost.

They called a hardware defense inside the memory itself the real long-term fix. Buyers cannot add that. What buyers can do is ask questions. Which mitigations does my cloud provider run on its NVIDIA GPUs? Does my own software trust the GPU by default? The speakers said that sanitizing everything the GPU sends to the CPU stops their CPU-side exploits.

The other side

This was a research demonstration. The excerpts describe no attacks in the wild, so the real-world scale is unclear. The excerpts also give no data behind the claim that each generation is worse. The speakers raised dedicated, unshared GPU clusters as an open question. They said they are still far from fully understanding the implications of GPU Rowhammer.

There is movement on defense. The researchers said they told NVIDIA in November 2025, then the major cloud providers. They said those providers are building their own mitigations.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from