GPU Glossary
/
Device Software·1 min read

Thread Block

The middle level of the CUDA thread hierarchy, a group of threads that run on one SM and can coordinate with each other.

A thread block is the level of the CUDA programming model's thread hierarchy between a grid and a thread. It is the abstract counterpart of the concrete cooperative thread arrays in PTX and SASS.

Thread blocks are an intermediate level of the thread group hierarchy of the CUDA programming model (left). A thread block executes on a single Streaming Multiprocessor (right, middle). Modified from diagrams in NVIDIA's CUDA Refresher: The CUDA Programming Model and the NVIDIA CUDA C++ Programming Guide.
Thread blocks are an intermediate level of the thread group hierarchy of the CUDA programming model (left). A thread block executes on a single Streaming Multiprocessor (right, middle). Modified from diagrams in NVIDIA's CUDA Refresher: The CUDA Programming Model and the NVIDIA CUDA C++ Programming Guide.

Blocks are the smallest unit of thread coordination the programming model exposes. A block executes on a single Streaming Multiprocessor.

Blocks must execute independently, so that any execution order is valid: fully serial in any order, or any interleaving.

Key idea Threads inside a block can cooperate; blocks themselves must not depend on each other's order.

Size#

A single CUDA kernel launch produces one or more thread blocks (as a thread block grid), and each block contains one or more warps. Blocks can be any size up to a limit of 1024 threads on current devices, but they are typically multiples of the warp size (32 on current devices).