图形/GPU 处理集群Graphics/GPU Processing Cluster
GPC 是一组 纹理处理集群 (TPC)（本身由 流式多处理器 (Streaming Multiprocessor)或 SM 组成）加上一个光栅引擎的集合。显然，有些人使用 NVIDIA GPU 进行图形处理，这时光栅引擎就很重要。相关地，该名称过去代表图形处理集群 (Graphics Processing Cluster)，但现在（例如在 NVIDIA CUDA C++编程指南 中）被扩展为"GPU 处理集群"。
自 H100 等 计算能力 (compute capability) 9.0 GPU 推出以来，CUDA 编程模型 (CUDA programming model) 的 线程层次结构 (thread hierarchy) 增加了一个额外层级，即调度到同一 GPC 上的 线程块 (thread block) "集群 (Cluster)" ，就像 线程块 (thread block) 的线程被调度到同一个 SM 上一样，且该层级拥有独立的 内存层次结构 (memory hierarchy)——分布式共享内存。在其他地方，我们略去了对此功能的讨论。
英文原文(Modal GPU Glossary)
A GPC is a collection of Texture Processing Clusters (TPCs) (themselves groups of Streaming Multiprocessors or SMs) plus a raster engine. Apparently, some people use NVIDIA GPUs for graphics, for which the raster engine is important. Relatedly, the name used to stand for Graphics Processing Cluster, but is now, e.g. in the NVIDIA CUDA C++ Programming Guide, expanded as "GPU Processing Cluster".
Since the introduction of compute capability 9.0 GPUs like H100s, there is an additional layer of the CUDA programming model's thread hierarchy, a "cluster" of thread blocks that are scheduled onto the same GPC, just as the threads of a thread block are scheduled onto the same SM, and have their own level of the memory hierarchy, distributed shared memory. Elsewhere, we elide discussion of this feature.
相关词条
本词条改编自 Modal GPU Glossary(CC BY 4.0)· 中文翻译 miter6/gpu-glossary-zh,MAE 整理排版。