A landfill scale-house now has two GPUs on the same box and a truck-ticket archive that barely fit on one device last season. How should they match the reader to scale?
Select an answer to reveal the explanation.
Short Explanation
Two GPUs on the same box, an archive that barely fit last season. Use single-GPU cuDF while the table fits, Dask cuDF when the read must split. Two devices are not an NCCL cluster, and a kernel does not pick the reader.
Full Explanation
Match the reader to scale: cuDF on one GPU while the table fits, Dask cuDF when the ingest must span GPUs. Two devices on the same box are not an invitation to NCCL multi-node design. Associate work does not write a kernel to choose the library. The decision is volume versus one device.