The problem
Can a device provide a bounded completion time while other work, memory behavior, and thermal state are changing?
What current research shows
GPU execution time depends on kernel configuration, memory transactions, dependencies, concurrent work, and device operating state. Throughput-oriented optimization can improve averages while leaving tail behavior difficult to bound.
Where the evidence stops
A real-time guarantee requires a defined device, workload, priority model, isolation policy, and bound. This study makes no universal guarantee.
What Valen Systems is testing
Define a bounded workload and report worst-case or high-percentile completion time, interference, thermal state, and recovery behavior.