GPU Module in KDB-X

This page explains what the GPU module is in KDB-X and when to use it.

If you would like to evaluate KDB-X GPU acceleration, please email devrel@kx.com with your request.

Key features

The GPU module:

• Supports moving data to and from the GPU with the ability to append data in batches.
• Supports a restricted set of qSQL queries on the GPU using standard q syntax.
• Supports sorting of in-memory tables on the GPU using standard q syntax.
• Supports as-of joins with attributes on the GPU using standard q syntax.

When to use the GPU module

Use the GPU module whenever you:

• Have workflows that benefit from parallelization across many threads.
• Have data resident in memory on the GPU and want improved performance.

Key terms

The following terms describe how data is stored, moved, and interpreted between the CPU and GPU:

• On-device: Data resident in GPU VRAM, represented in q as foreign.
• On-host: Data resident in CPU RAM, accessible to standard q operations.
• Foreign: The q representation of an on-device object.
• Mixed residency: A table where some columns are on-device and others on-host.
• Grouped attribute: A q attribute (`g#) required for certain GPU operations such as .gpu.bin and .gpu.aj.
• VRAM: GPU memory, distinct from system RAM.
• CUDA kernel: A function executed in parallel across GPU threads, distinct from the OS kernel.

Next steps

• To get started with the GPU module, refer to the Quickstart.