Scope
This project investigates the performance characteristics of OLAP (Online Analytical Processing) database systems on the NVIDIA GH200 platform, a heterogeneous architecture that tightly integrates Grace CPUs and Hopper GPUs via a high-bandwidth, low-latency interconnect. Traditional OLAP database systems are primarily optimized for CPU-centric execution, but the GH200 introduces new opportunities for accelerating analytical workloads by exploiting both massive GPU parallelism and fast, unified memory access across CPU and GPU domains. Traditional approaches are typically limited by the available GPU memory, and scaling beyond this creates a massive data movement bottleneck due to the slow PCIe interface. Due to the nature of the GH200 system, there is a fast interconnect between CPU and GPU memory, which enables faster data transfer between the CPU and the GPU.
