MaxBench: Benchmarking GPU Interconnect Impact on Relational Data Analytics
Researchers introduce MaxBench, a comprehensive framework for benchmarking and profiling relational data analytics workloads on GPUs. It evaluates the performance impact of various GPU models (RTX3090, A100, H100, Grace Hopper GH200) and interconnects (PCIe 3.0, 4.0, 5.0, and NVLink 4.0) on workloads like TPC-H, H2O-G, and ClickBench. Moving beyond traditional metrics like arithmetic intensity and GFlop/s, MaxBench proposes 'characteristic query complexity' and 'characteristic GPU efficiency' and uses a novel cost model to predict query execution performance. The study reveals trade-offs between GPU compute capacity and interconnect bandwidth and uses the model to project the impact of future interconnect bandwidth or GPU efficiency improvements.
Read more