Texture Processing ClusterTPC
A pair of adjacent Streaming Multiprocessors, which since Blackwell maps onto the "CTA pair" level of the PTX thread hierarchy.
A Texture Processing Cluster (TPC) is a pair of adjacent Streaming Multiprocessors (SMs).
Before the Blackwell SM architecture, TPCs did not map onto any level of the CUDA programming model's memory hierarchy or thread hierarchy.
CTA pairs in Blackwell#
Blackwell's fifth-generation Tensor Cores added a "CTA pair" level to the Parallel Thread eXecution (PTX) thread hierarchy, and it maps onto TPCs.
Many tcgen05 PTX instructions have a .cta_group field that selects either:
- a single SM (
.cta_group::1), or - the pair of SMs in a TPC (
::2).
These map to the 1SM and 2SM variants of Streaming Assembler (SASS) instructions like MMA.
Related#
- Graphics/GPU Processing Cluster: the next level up, a group of TPCs
- Cooperative Thread Array: the "CTA" in CTA pair
- Thread Hierarchy: where CTA pairs fit in