4 ms·
Confidentiality Issues on a GPU in a Virtualized Environment [pdf]
- lmeyerov 12y agoI've been thinking increasingly that webcl's place is in the cloud, including because of all the work on multiuser/multitenant confidentiality+integrity+availability!
- j_s 12y agoTL;DR: We found that the GPU global memory is zeroed only in some configurations.
- maccard 12y agoNoticed this with CUDA before. When I tried to render data that lay outside the buffers I was working on, I would get textures of my web browser appearing instead of what I expected.
- iamsalman 12y agoCurrently, only AWS offers GPU instances. There are other, smaller providers like Nimbix or Peer1 but they offer bare-metal instances. With AWS, they currently have Grid K520 GPUs which is targeted for public and private Clouds. The gist of this paper is that non ECC memory does not zero the memory but it conveniently skips the fact that public or private Cloud uses either NVIDIA Tesla or Quadro series cards, both of which have ECC (one of the main reasons they're expensive than say GTX Titan which offers same single/double-precision performance but does not offer ECC). No one really uses GeForce series cards for anything serious and definitely not virtualization so even though this is a vulnerability, it may not effect more than 1e-10% of virtualized GPUs out there