What happens in the GPU doesn't stay in the GPU

This talk explores what it means when containers access GPUs in a cluster and the risks for multi-tenant environments, including whether sensitive data could leak between tenants. GPUs were not designed for cloud environments where many applications share the same hardware. The session examines GPU access for containers, including hardware versus software isolation, how GPU memory works, and how GPU vendor lock-in can affect security posture. It also covers GPU passthrough, vGPUs, GPU exploits, and application isolation.