Powerful GPUs or Fast Interconnects: Analyzing Relational Workloads
16 days ago
- #GPU Performance
- #Interconnect Bandwidth
- #Data Analytics
- Study explores impact of GPU models (RTX3090, A100, H100, GH200) and interconnects (PCIe 3.0-5.0, NVLink 4.0) on relational workloads (TPC-H, H2O-G, ClickBench).
- Introduces MaxBench, a framework for benchmarking, profiling, and modeling GPU workloads, including a novel cost model for query execution performance.
- Proposes characteristic query complexity and characteristic GPU efficiency as new metrics for data analytics workloads, moving beyond traditional metrics.
- Conducts extensive experimental analysis to reveal trade-offs between GPU computing capacity and interconnect bandwidth.
- Uses cost model to examine future performance trends based on improvements in interconnect bandwidth or GPU efficiency.