GPU Glossary
/
Device Hardware·1 min read

Graphics/GPU Processing ClusterGPC

A collection of Texture Processing Clusters plus a raster engine; since compute capability 9.0, thread block clusters are scheduled onto one GPC.

A GPC is a collection of Texture Processing Clusters (TPCs), which are themselves groups of Streaming Multiprocessors (SMs), plus a raster engine.

The raster engine matters for graphics, which some people still use NVIDIA GPUs for. Relatedly, GPC used to stand for Graphics Processing Cluster. It is now expanded as "GPU Processing Cluster", e.g. in the NVIDIA CUDA C++ Programming Guide.

Clusters in the programming model#

Compute capability 9.0 GPUs like the H100 added a layer to the CUDA programming model's thread hierarchy: a "cluster" of thread blocks.

  • A cluster's thread blocks are scheduled onto the same GPC, just as a thread block's threads are scheduled onto the same SM.
  • Clusters get their own level of the memory hierarchy: distributed shared memory.

The rest of this glossary does not cover this feature.