AWS and Qumulo test AI training across cloud regions without copying data

AWS and Qumulo say a SageMaker HyperPod training cluster in Oregon matched the throughput of a cluster beside its dataset in Ohio after an initial warmup. In their test, both ran the same 1.02 billion-parameter model; the remote cluster reached 115-117 samples per second while reading data across a link with 60 ms of latency. Qumulo's software presents the remote dataset through a local file mount and predicts which blocks to cache near the GPUs. The result could let training teams use compute in another region without copying an entire dataset. The remote run was 15-20 percent slower for its first 100-150 batches.