A graphics processing unit kernel describes computation that runs concurrently across a grid of threads. Each thread handles part of the data, while the kernel's indexing, memory access, synchronization, and launch configuration determine how work maps onto the hardware.
Fast kernels reduce memory traffic, avoid unnecessary synchronization, keep execution units occupied, and use the appropriate memory hierarchy. Multi-device kernels add data partitioning and communication, making correctness and performance sensitive to ordering and transfer choices.
Acronyms and aliases
accelerator kernel variantGPU kernel variant
Related terms
Frequently asked questions
What does a graphics processing unit kernel do?
It applies a numerical operation across many data elements in parallel using threads scheduled on the graphics processor.
Why can two correct graphics processing unit kernels have different speeds?
They may use memory, synchronization, thread scheduling, communication, and hardware resources with very different efficiency.
