GPU Architecture, Environments, and Performance
Free NVIDIA-Certified Associate: Accelerated Data Science practice — 6 questions on GPU Architecture, Environments, and Performance, with explanations. No sign-up. Full 12-question mixed test →
A data scientist runs a cuDF ETL pipeline on a single A100 GPU with 40GB of memory. A groupby-join on a 60GB dataset throws a CUDA out-of-memory error even though RMM's pool allocator is already enabled. Which change is MOST likely to resolve this without adding more GPUs?
When a dataset exceeds physical GPU memory, dask-cuda's out-of-core spilling (or Unified Memory oversubscription) lets data that doesn't fit spill to host memory automatically, letting the join complete instead of crashing.
A team scales a cuML training job across 8 GPUs connected via NVLink on a single DGX node using dask-cuda. Which communication configuration should they enable to maximize GPU-to-GPU transfer throughput between workers?
dask-cuda is designed to use UCX, which automatically selects the fastest available transport — NVLink for intra-node GPU-to-GPU transfers and InfiniBand/RDMA for inter-node — giving far higher bandwidth than generic networking stacks.
An engineer suspects a custom CUDA kernel invoked inside a cuML estimator has low occupancy and high register pressure. Which NVIDIA tool provides kernel-level metrics such as achieved occupancy and registers-per-thread to diagnose this?
Nsight Compute is the kernel-level profiler that reports detailed per-kernel metrics — occupancy, register usage, memory throughput, and warp stall reasons — exactly what's needed to diagnose kernel-level bottlenecks.
When initializing RMM's pool allocator for a RAPIDS workflow that performs many repeated small allocations, which configuration approach is recommended to minimize fragmentation and allocation overhead?
Pre-reserving a large upfront pool (well above the smallest workload needs) lets RMM sub-allocate from a single large device buffer, avoiding repeated cudaMalloc/cudaFree calls and reducing fragmentation from many small allocations.
Why do data-parallel workloads such as cuDF dataframe operations achieve large speedups on GPUs compared to CPUs?
GPUs pair high-bandwidth HBM memory (on the order of terabytes/sec versus tens of GB/s for typical CPU DRAM channels) with thousands of parallel cores, so the same column/element-wise operation can be applied to massive amounts of data simultaneously — the core reason data-parallel operations like cuDF column transforms scale so well.
A company is deploying a multi-node RAPIDS cluster on cloud GPU instances and must guarantee every worker node runs an identical, driver-compatible software stack with minimal setup time per node. Which environment strategy BEST meets this requirement?
Pulling the same pre-built, version-pinned RAPIDS container from NGC on every node guarantees an identical CUDA/RAPIDS/driver-compatible stack with fast, repeatable deployment — the standard recommended approach for reproducible multi-node RAPIDS clusters.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$99.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →