GPUHammer Strikes: How One Bit Flip Can Turn Your AI Dreams into Nightmares
The University of Toronto researchers have taken the Rowhammer attack to new heights—or rather, pixels—by targeting GPUs with the aptly named GPUHammer. This attack on an Nvidia GPU sent machine learning models into accuracy freefall, dropping from 80% to 0.1%. Turns out, hammering isn’t just for nails anymore; it’s perfect for shaking up GPUs too!
Hot Take:
Brace yourselves, folks! The GPU apocalypse is nigh! The University of Toronto researchers have found a way to give GPUs a taste of the Rowhammer attack, and it’s practically hammer time for machine learning models. Who knew that the humble bit flip could be such a party pooper?
Key Points:
- Rowhammer attacks have now expanded their horizons from CPUs to GPUs, thanks to the University of Toronto researchers.
- GPUHammer attack was successfully demonstrated on an NVIDIA A6000 GPU, impacting machine learning models.
- A single bit flip could drastically drop a model’s accuracy from 80% to a meager 0.1%.
- Nvidia confirmed the vulnerability and suggested ECC as a mitigation, albeit with performance trade-offs.
- Testing other GPUs is challenging due to the soldered DRAM modules and the high cost of GPUs.
Already a member? Log in here