“repeated boot and stress testing shows stable GPU enumeration” addresses card retention and connector seating were not completed to the platform standard, but the current hardware or infrastructure issue is physical reassembly and cooling acceptance were skipped. The option “thermal stress run and hardware inspections meet post-service acceptance” is more directly tied to the required outcome.

“chassis fit check and fan telemetry satisfy service specifications” addresses physical compatibility and airflow clearance were not confirmed, but the current hardware or infrastructure issue is physical reassembly and cooling acceptance were skipped. The option “thermal stress run and hardware inspections meet post-service acceptance” is more directly tied to the required outcome.

“board power status is healthy and full-load draw stays within limits” addresses the GPU power connection and supply budget are incomplete, but the current hardware or infrastructure issue is physical reassembly and cooling acceptance were skipped. The option “thermal stress run and hardware inspections meet post-service acceptance” is more directly tied to the required outcome.

“run supported concurrent GPU stress and interconnect health checks” addresses multi-device concurrency and topology were not validated, but the current hardware or infrastructure issue is device enumeration is not sufficient evidence of functional compute readiness. The option “execute supported CUDA kernel tests and investigate runtime errors” is more directly tied to the required outcome.

“inspect per-rank memory placement and GPU resource partition assignments” addresses per-device allocation and partitioning do not match application memory requirements, but the current hardware or infrastructure issue is device enumeration is not sufficient evidence of functional compute readiness. The option “execute supported CUDA kernel tests and investigate runtime errors” is more directly tied to the required outcome.

Useful for intra-node collective communication path is not delivering expected bandwidth, but it does not remove the specific condition: device enumeration is not sufficient evidence of functional compute readiness.

“an eight-GPU test completes with no Xid errors or missing peers” addresses multi-device concurrency and topology were not validated, but the current hardware or infrastructure issue is device enumeration is not sufficient evidence of functional compute readiness. The option “bounded compute sample completes with correct results and no device error” is more directly tied to the required outcome.

“each process fits its assigned GPU and a representative batch completes” addresses per-device allocation and partitioning do not match application memory requirements, but the current hardware or infrastructure issue is device enumeration is not sufficient evidence of functional compute readiness. The option “bounded compute sample completes with correct results and no device error” is more directly tied to the required outcome.

“representative kernels complete and outputs satisfy numerical tolerance” addresses application binaries were not validated against installed GPU compute capability, but the current hardware or infrastructure issue is device enumeration is not sufficient evidence of functional compute readiness. The option “bounded compute sample completes with correct results and no device error” is more directly tied to the required outcome.

Useful for application binaries were not validated against installed GPU compute capability, but it does not remove the specific condition: local GPU interconnect bandwidth is degraded or uninitialized.

“run NCCL topology-sensitive benchmarks and isolate the failing local interconnect” addresses intra-node collective communication path is not delivering expected bandwidth, but the current hardware or infrastructure issue is local GPU interconnect bandwidth is degraded or uninitialized. The option “verify NVLink paths and peer-to-peer transfer topology” is more directly tied to the required outcome.

“profile CPU I/O preprocessing and feeding of GPU kernels” addresses the input pipeline starves accelerators of batches, but the current hardware or infrastructure issue is local GPU interconnect bandwidth is degraded or uninitialized. The option “verify NVLink paths and peer-to-peer transfer topology” is more directly tied to the required outcome.

“NCCL output meets qualified intra-node bandwidth for the GPU count” addresses intra-node collective communication path is not delivering expected bandwidth, but the current hardware or infrastructure issue is local GPU interconnect bandwidth is degraded or uninitialized. The option “measured peer bandwidth matches the qualified platform acceptance floor” is more directly tied to the required outcome.

“end-to-end batch latency improves with measured transfer utilization” addresses memory movement and transfer path constrain end-to-end application throughput, but the current hardware or infrastructure issue is local GPU interconnect bandwidth is degraded or uninitialized. The option “measured peer bandwidth matches the qualified platform acceptance floor” is more directly tied to the required outcome.

“bounded compute sample completes with correct results and no device error” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is local GPU interconnect bandwidth is degraded or uninitialized. The option “measured peer bandwidth matches the qualified platform acceptance floor” is more directly tied to the required outcome.

“inspect throttling reasons power limits and authorized clock settings” addresses device performance is constrained by power or application clock policy, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “inspect per-rank memory placement and GPU resource partition assignments” is more directly tied to the required outcome.

“profile pinned-memory transfers and optimize host-device staging within policy” addresses memory movement and transfer path constrain end-to-end application throughput, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “inspect per-rank memory placement and GPU resource partition assignments” is more directly tied to the required outcome.

“run supported concurrent GPU stress and interconnect health checks” addresses multi-device concurrency and topology were not validated, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “inspect per-rank memory placement and GPU resource partition assignments” is more directly tied to the required outcome.

“representative kernels complete and outputs satisfy numerical tolerance” addresses application binaries were not validated against installed GPU compute capability, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “each process fits its assigned GPU and a representative batch completes” is more directly tied to the required outcome.

“end-to-end batch latency improves with measured transfer utilization” addresses memory movement and transfer path constrain end-to-end application throughput, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “each process fits its assigned GPU and a representative batch completes” is more directly tied to the required outcome.

“bounded compute sample completes with correct results and no device error” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is per-device allocation and partitioning do not match application memory requirements. The option “each process fits its assigned GPU and a representative batch completes” is more directly tied to the required outcome.

Useful for memory movement and transfer path constrain end-to-end application throughput, but it does not remove the specific condition: the input pipeline starves accelerators of batches.

“verify NVLink paths and peer-to-peer transfer topology” addresses local GPU interconnect bandwidth is degraded or uninitialized, but the current hardware or infrastructure issue is the input pipeline starves accelerators of batches. The option “profile CPU I/O preprocessing and feeding of GPU kernels” is more directly tied to the required outcome.

“execute supported CUDA kernel tests and investigate runtime errors” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is the input pipeline starves accelerators of batches. The option “profile CPU I/O preprocessing and feeding of GPU kernels” is more directly tied to the required outcome.

“an eight-GPU test completes with no Xid errors or missing peers” addresses multi-device concurrency and topology were not validated, but the current hardware or infrastructure issue is the input pipeline starves accelerators of batches. The option “improved data loader throughput coincides with sustained GPU occupancy” is more directly tied to the required outcome.

“end-to-end batch latency improves with measured transfer utilization” addresses memory movement and transfer path constrain end-to-end application throughput, but the current hardware or infrastructure issue is the input pipeline starves accelerators of batches. The option “improved data loader throughput coincides with sustained GPU occupancy” is more directly tied to the required outcome.

“bounded compute sample completes with correct results and no device error” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is the input pipeline starves accelerators of batches. The option “improved data loader throughput coincides with sustained GPU occupancy” is more directly tied to the required outcome.

“run NCCL topology-sensitive benchmarks and isolate the failing local interconnect” addresses intra-node collective communication path is not delivering expected bandwidth, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “run supported concurrent GPU stress and interconnect health checks” is more directly tied to the required outcome.

“build and verify a compatible CUDA runtime and kernel target” addresses application binaries were not validated against installed GPU compute capability, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “run supported concurrent GPU stress and interconnect health checks” is more directly tied to the required outcome.

“execute supported CUDA kernel tests and investigate runtime errors” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “run supported concurrent GPU stress and interconnect health checks” is more directly tied to the required outcome.

“bounded compute sample completes with correct results and no device error” addresses device enumeration is not sufficient evidence of functional compute readiness, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “an eight-GPU test completes with no Xid errors or missing peers” is more directly tied to the required outcome.

“each process fits its assigned GPU and a representative batch completes” addresses per-device allocation and partitioning do not match application memory requirements, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “an eight-GPU test completes with no Xid errors or missing peers” is more directly tied to the required outcome.

“end-to-end batch latency improves with measured transfer utilization” addresses memory movement and transfer path constrain end-to-end application throughput, but the current hardware or infrastructure issue is multi-device concurrency and topology were not validated. The option “an eight-GPU test completes with no Xid errors or missing peers” is more directly tied to the required outcome.