Understand Where GPU Acceleration Delivers The Most Value
Large time-series workloads often combine operations with very different execution characteristics. Joins, sorts, and windowed aggregations can map efficiently to thousands of parallel GPU threads. Other operations may remain faster on a CPU, particularly when they depend on index shortcuts or require frequent transfers between host and device memory.
This whitepaper examines both sides of that performance equation. It explains how the KDB-X GPU module works, where it delivers substantial speedups today, and why workload shape and data residency determine the result.
What The Whitepaper Covers
- How the GPU module integrates CUDA with KDB-X through the C API
- How to move data between CPU and GPU memory
- How KDB-X executes qSQL-style queries against GPU-resident tables
- The GPU execution model behind joins, sorts, reductions, sliding
windows, and filters - Benchmark results from the KX NYSE TAQ query suite
- Where GPU execution outperforms the CPU today
- How transfer costs and current indexing limitations affect performance
- How performance scales from one GPU to four GPUs
