Speculative Execution on Multi-GPU Systems