NCP-AII 試験問題を無料オンラインアクセス

What command is needed to measure BER (Bit Error Rate)?

• A.ethtool -S < device >
• B.mlxconfig -d < device > q
• C.mstflint -d < device > q full
• D.mlxlink -d < device > -c -e

Your tasked with updating both NVIDIA GPU drivers and DOCA drivers on a set of servers used for AI workloads. The environment previously had an older driver stack and custom kernel modules. What is the most important step to successfully upgrade the drivers without causing conflicts?

• A.Keep the older driver running alongside the new version in case you need to roll back the upgrade.
• B.Validate the driver version post-install since the fresh install will overwrite the legacy drivers.
• C.Update the GPU driver leaving the DOCA and OFED drivers unchanged as long as they are detecting the hardware properly.
• D.Uninstall all existing GPU and DOCA-related drivers and associated kernel modules before the new install.

During a multi-day NeMo burn-in, intermittent " GPU fell off bus " errors occur. Which diagnostic approach isolates hardware faults?

• A.Run DCGM diagnostics alongside burn-in to monitor GPU health metrics.
• B.Reduce blocksize to 500MB to lower memory pressure.
• C.Enable HPL_USE_NVSHMEM for alternative memory sharing.
• D.Switch from BERT to GPT models for simpler computations.

You are standing up an NVIDIA DGX system for enterprise production. Stakeholder teams require system reliability, performance consistency under load, and proper escalation processes before release. A recent system in another cluster experienced intermittent GPU failures attributed to missed early-stage validation.
Which deployment and validation sequence best addresses production readiness and mitigates the risk of avoidable downtime or performance loss?

• A.Update network topology, assign static IPs and DNS entries, register the system with NVIDIA, then conduct basic OS-level checks and enable user access after login testing is successful.
• B.Install latest OS images and drivers, confirm OS and container functionality, invite users for a monitored production trial, and collect workload feedback to plan any further diagnostics or updates.
• C.Complete hardware and cabling, power on the system, update firmware and drivers, run full hardware health checks and stress diagnostics using NVSM, verify all GPU and system sensor logs, and validate GPU accessibility.
• D.Power on the system, install all AI frameworks, configure the CUDA and library stack, set up user environments, then plan stress tests and diagnostics as part of ongoing routine operations.

During HPL execution on a DGX cluster, the benchmark fails with " not enough memory " errors despite sufficient physical RAM. Which HPL.dat parameter adjustment is most effective?

• A.Reduce the problem size while maintaining the same block size.
• B.Increase block size to 6144 to maximize GPU utilization.
• C.Set PMAP to 1 to enable process mapping.
• D.Disable double-buffering via BCAST parameter.