GPU clusters need synthetic stress tests to prevent AI training failures

1. Run synthetic stress tests across all GPUs simultaneously rather than relying on individual node health checks.
2. Benchmark inter-node network communication bandwidth under maximum load to identify hidden bottlenecks.
3. Validate memory bandwidth and thermal stability by running sustained mock training workloads before launching production jobs.