GPU timing side-channels for workload inference

This post contributes to the understanding of security implications in shared GPU environments by detailing a timing side-channel attack. The attack, which probes only the attacker's own memory, can infer model depth, decoding speed, and distinguish workload categories (code, prose, JSON, retrieval) by analyzing timing signals. This highlights that memory and compute isolation alone do not eliminate hardware-level signals that can leak information about co-located tenants. The work emphasizes the need for careful tenant placement and understanding isolation tradeoffs.