设备硬件 · TPC

纹理处理集群Texture Processing Cluster

纹理处理集群 (Texture Processing Cluster, TPC) 是一对相邻的流式多处理器 (Streaming Multiprocessors, SMs)。

在 Blackwell SM 架构之前，TPC 并未映射到 CUDA 编程模型的内存层次结构或线程层次结构的任何层级。

Blackwell SM 架构中的第五代 Tensor Cores 在并行线程执行 (Parallel Thread eXecution, PTX) 的线程层次结构中增加了 "CTA 对" 层级，该层级映射到 TPC。许多 tcgen05 PTX 指令包含一个 .cta_group 字段，可以使用单个流式多处理器 (.cta_group::1) 或 TPC 中的一对流式多处理器 (::2)，并分别对应 MMA 等流式汇编器 (Streaming Assembler, SASS) 指令（如 MMA）的 1SM 和 2SM 变体。

英文原文(Modal GPU Glossary)

A Texture Processing Cluster (TPC) is a pair of adjacent Streaming Multiprocessors (SMs).

Before the Blackwell SM architecture, TPCs were not mapped onto any level of the CUDA programming model's memory hierarchy or thread hierarchy.

The fifth-generation Tensor Cores in the Blackwell SM architecture added the "CTA pair" level of the Parallel Thread eXecution (PTX) thread hierarchy, which maps onto TPCs. Many tcgen05 PTX instructions include a .cta_group field that can use a single SM (.cta_group::1) or a pair of SMs in a TPC (::2), which are mapped to 1SM and 2SM variants of Streaming Assembler (SASS) instructions like MMA.

相关词条