“sustained workload metrics show no unexpected GPU throttling flags” addresses full installed-GPU visibility does not prove healthy load behavior, but the current hardware or infrastructure issue is device persistence settings do not match the intended repeated-job operating profile. The option “repeated short-job startup measurements show reduced initialization overhead” is more directly tied to the required outcome.

“nvidia-smi lists the intended GPU count with no driver initialization errors” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is device persistence settings do not match the intended repeated-job operating profile. The option “repeated short-job startup measurements show reduced initialization overhead” is more directly tied to the required outcome.

“nvidia-smi and platform logs show all eight devices healthy” addresses a physical or driver-level enumeration fault affects one accelerator, but the current hardware or infrastructure issue is device persistence settings do not match the intended repeated-job operating profile. The option “repeated short-job startup measurements show reduced initialization overhead” is more directly tied to the required outcome.

“verify the supported driver kernel modules and device enumeration” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “validate installed CUDA runtime and application compatibility against the driver” is more directly tied to the required outcome.

“inspect GPU memory and compute capability and run the approved workload smoke test” addresses scheduling readiness did not confirm required device memory and capabilities, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “validate installed CUDA runtime and application compatibility against the driver” is more directly tied to the required outcome.

“investigate GPU throttle reasons and restore thermal or power headroom” addresses full installed-GPU visibility does not prove healthy load behavior, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “validate installed CUDA runtime and application compatibility against the driver” is more directly tied to the required outcome.

“an inventory sample maps every reported GPU ID to the expected physical slot” addresses accelerator identity lacks a stable mapping to the physical PCIe slot, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “a sample CUDA application initializes and reports the expected device” is more directly tied to the required outcome.

“device inventory meets requirements and the test allocates memory without failure” addresses scheduling readiness did not confirm required device memory and capabilities, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “a sample CUDA application initializes and reports the expected device” is more directly tied to the required outcome.

“topology display and local bandwidth checks match the approved HGX design” addresses interconnect readiness is not captured in GPU bring-up acceptance, but the current hardware or infrastructure issue is driver visibility was mistaken for full application-runtime readiness. The option “a sample CUDA application initializes and reports the expected device” is more directly tied to the required outcome.

“check GPU seating device power and driver health before re-adding node to scheduling” addresses a physical or driver-level enumeration fault affects one accelerator, but the current hardware or infrastructure issue is full installed-GPU visibility does not prove healthy load behavior. The option “investigate GPU throttle reasons and restore thermal or power headroom” is more directly tied to the required outcome.

“verify the supported driver kernel modules and device enumeration” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is full installed-GPU visibility does not prove healthy load behavior. The option “investigate GPU throttle reasons and restore thermal or power headroom” is more directly tied to the required outcome.

Useful for interconnect readiness is not captured in GPU bring-up acceptance, but it does not remove the specific condition: full installed-GPU visibility does not prove healthy load behavior.

“repeated short-job startup measurements show reduced initialization overhead” addresses device persistence settings do not match the intended repeated-job operating profile, but the current hardware or infrastructure issue is full installed-GPU visibility does not prove healthy load behavior. The option “sustained workload metrics show no unexpected GPU throttling flags” is more directly tied to the required outcome.

“utilization rises during the job and diagnostic compute completes successfully” addresses an idle utilization sample does not test compute functionality, but the current hardware or infrastructure issue is full installed-GPU visibility does not prove healthy load behavior. The option “sustained workload metrics show no unexpected GPU throttling flags” is more directly tied to the required outcome.

“nvidia-smi lists the intended GPU count with no driver initialization errors” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is full installed-GPU visibility does not prove healthy load behavior. The option “sustained workload metrics show no unexpected GPU throttling flags” is more directly tied to the required outcome.

Useful for accelerator identity lacks a stable mapping to the physical PCIe slot, but it does not remove the specific condition: ephemeral GPU index positions are unsuitable as maintenance identities.

“verify the supported driver kernel modules and device enumeration” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is ephemeral GPU index positions are unsuitable as maintenance identities. The option “record device UUID and PCI identifiers instead of relying on index alone” is more directly tied to the required outcome.

“check GPU seating device power and driver health before re-adding node to scheduling” addresses a physical or driver-level enumeration fault affects one accelerator, but the current hardware or infrastructure issue is ephemeral GPU index positions are unsuitable as maintenance identities. The option “record device UUID and PCI identifiers instead of relying on index alone” is more directly tied to the required outcome.

“an inventory sample maps every reported GPU ID to the expected physical slot” addresses accelerator identity lacks a stable mapping to the physical PCIe slot, but the current hardware or infrastructure issue is ephemeral GPU index positions are unsuitable as maintenance identities. The option “UUID-based inventory remains stable through a controlled reboot” is more directly tied to the required outcome.

“device inventory meets requirements and the test allocates memory without failure” addresses scheduling readiness did not confirm required device memory and capabilities, but the current hardware or infrastructure issue is ephemeral GPU index positions are unsuitable as maintenance identities. The option “UUID-based inventory remains stable through a controlled reboot” is more directly tied to the required outcome.

“nvidia-smi lists the intended GPU count with no driver initialization errors” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is ephemeral GPU index positions are unsuitable as maintenance identities. The option “UUID-based inventory remains stable through a controlled reboot” is more directly tied to the required outcome.

“verify the supported driver kernel modules and device enumeration” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is scheduling readiness did not confirm required device memory and capabilities. The option “inspect GPU memory and compute capability and run the approved workload smoke test” is more directly tied to the required outcome.

“validate installed CUDA runtime and application compatibility against the driver” addresses driver visibility was mistaken for full application-runtime readiness, but the current hardware or infrastructure issue is scheduling readiness did not confirm required device memory and capabilities. The option “inspect GPU memory and compute capability and run the approved workload smoke test” is more directly tied to the required outcome.

Useful for interconnect readiness is not captured in GPU bring-up acceptance, but it does not remove the specific condition: scheduling readiness did not confirm required device memory and capabilities.

“topology display and local bandwidth checks match the approved HGX design” addresses interconnect readiness is not captured in GPU bring-up acceptance, but the current hardware or infrastructure issue is scheduling readiness did not confirm required device memory and capabilities. The option “device inventory meets requirements and the test allocates memory without failure” is more directly tied to the required outcome.

“utilization rises during the job and diagnostic compute completes successfully” addresses an idle utilization sample does not test compute functionality, but the current hardware or infrastructure issue is scheduling readiness did not confirm required device memory and capabilities. The option “device inventory meets requirements and the test allocates memory without failure” is more directly tied to the required outcome.

“an inventory sample maps every reported GPU ID to the expected physical slot” addresses accelerator identity lacks a stable mapping to the physical PCIe slot, but the current hardware or infrastructure issue is scheduling readiness did not confirm required device memory and capabilities. The option “device inventory meets requirements and the test allocates memory without failure” is more directly tied to the required outcome.

Useful for host driver registration has not established visibility of installed GPUs, but it does not remove the specific condition: interconnect readiness is not captured in GPU bring-up acceptance.

“inspect GPU memory and compute capability and run the approved workload smoke test” addresses scheduling readiness did not confirm required device memory and capabilities, but the current hardware or infrastructure issue is interconnect readiness is not captured in GPU bring-up acceptance. The option “query supported GPU interconnect topology and verify expected links” is more directly tied to the required outcome.

“investigate GPU throttle reasons and restore thermal or power headroom” addresses full installed-GPU visibility does not prove healthy load behavior, but the current hardware or infrastructure issue is interconnect readiness is not captured in GPU bring-up acceptance. The option “query supported GPU interconnect topology and verify expected links” is more directly tied to the required outcome.

“device inventory meets requirements and the test allocates memory without failure” addresses scheduling readiness did not confirm required device memory and capabilities, but the current hardware or infrastructure issue is interconnect readiness is not captured in GPU bring-up acceptance. The option “topology display and local bandwidth checks match the approved HGX design” is more directly tied to the required outcome.

“a sample CUDA application initializes and reports the expected device” addresses driver visibility was mistaken for full application-runtime readiness, but the current hardware or infrastructure issue is interconnect readiness is not captured in GPU bring-up acceptance. The option “topology display and local bandwidth checks match the approved HGX design” is more directly tied to the required outcome.

“nvidia-smi lists the intended GPU count with no driver initialization errors” addresses host driver registration has not established visibility of installed GPUs, but the current hardware or infrastructure issue is interconnect readiness is not captured in GPU bring-up acceptance. The option “topology display and local bandwidth checks match the approved HGX design” is more directly tied to the required outcome.