Ada To Hopper, GDDR6 Is Potentially Vulnerable To Rowhammer
You can either thank or curse the researchers at the University of Toronto for discovering that Rowhammer isn’t just for DRAM anymore. They provided a proof of concept attack against NVIDIA A6000 GPUs using GDDR6, which spans the vast majority of their HPC cards. The good news first, seeing as how it is Friday, is that the attack only works when System-Level ECC isn’t enabled. You would think that if you are developing LLMs and other HPC workloads, that you would want error correction enabled to ensure the integrity of the data, let alone for security, however that is not always the case.
The reason is that enabling ECC does have an impact on the speed workloads complete in. The amount of impact on the performance varies depending on what you are doing, with some reporting up to a 50% decrease in the speed at which their workloads complete. This is an extreme example, but even a 10% increase does translate to higher operating costs. There are also features on older cards that aren’t compatible with ECC, though again they are few and far between.
One solution is to migrate to newer Hopper or Blackwell cards, which come with built-in on-die ECC protection enabled by default. You can choose to turn that feature off if you are exceptionally cheap, but that means Rowhammer attacks are possible against your hardware. Bleeping Computer lists the cards which are vulnerable as well as those which ship with the protections already enabled.
Source link