Graphics processing units (GPUs) can make Big Data analytics 100 times faster than systems based entirely on Central Processing Units (CPUs). OmniSci harnesses the parallel processing power of GPUs to deliver the world’s fastest data queries and visualization.
However, most analytic workflows still move data between CPUs and GPUs, and so much of the speedup provided by a GPU can be lost during transfers to a CPU. NVIDIA created NVIDIA® NVLink™ to reduce that bottleneck, and it is only available for CPU-to-GPU transfers on IBM POWER systems.
This whitepaper provides OmniSci performance benchmark results on IBM Power System S822LC for High Performance Computing and shows how NVLink speeds data science workloads shared across NVIDIA® Tesla® P100 GPUs and IBM POWER8 CPUs.