设备硬件

GPU 核心GPU Core

核心是构成流式多处理器 (SMs) 的主要计算单元。

H100 GPU 流式多处理器的内部架构示意图。图中绿色部分为 GPU 核心，褐红色部分为其他计算单元，橙色部分为调度单元，蓝色部分为内存。修改自 NVIDIA 的 H100 白皮书。

GPU 核心类型的举例包括 CUDA 核心和 Tensor 核心。

尽管 GPU 核心与 CPU 核心在 “执行实际计算的组件” 这一点上具有可比性，但这种类比可能存在较大误导性。或许，从量化计算机架构师的角度出发，将它们视为数据流入并返回转换后数据的"管道"可能更有帮助。从硬件角度来看，这些管道与特定的指令相关联；从程序员角度来看，则对应不同的基础吞吐量特性（例如 Tensor 核心的浮点矩阵乘法算术吞吐量）。

相比之下，流式多处理器 (SMs) 更接近于 CPU 核心的等效概念，因为它们拥有存储信息的寄存器内存、用于转换数据的核心，以及用于指定和控制转换操作的指令调度器。

英文原文(Modal GPU Glossary)

The cores are the primary compute units that make up the Streaming Multiprocessors (SMs).

.](themed-image://gh100-sm.svg)

Examples of GPU core types include CUDA Cores and Tensor Cores.

Though GPU cores are comparable to CPU cores in that they are the component that effects actual computations, this analogy can be quite misleading. Instead, it is perhaps more helpful to take the viewpoint of the quantitative computer architect and think of them as "pipes" into which data goes in and out of which transformed data is returned. These pipes are associated in turn with specific instructions from the hardware's perspective and with different fundamental affordances of throughput from the programmers' (e.g. floating point matrix multiplication arithmetic throughput in the case of the Tensor Cores).

The SMs are closer to being the equivalent of CPU cores, in that they have register memory to store information, cores to transform it, and an instruction scheduler to specify and command transformations.

相关词条