GPU 如何颠覆高性能计算

GPU 如何颠覆高性能计算
How GPUs Are Disrupting High-Performance Computing

原始链接: https://www.zerohedge.com/technology/how-gpus-are-disrupting-high-performance-computing

图形处理单元 (GPU) 最初是为渲染视频游戏图形而设计的，现在在高性能计算 (HPC) 中发挥着重要作用。他们改变了从人工智能 (AI) 训练到浮游动物分类等任务。 CPU、传统中央处理单元和 GPU 均由负责执行指令的计算元件、管理操作的控制元件和内存系统组成。然而，它们的结构差异很大。 CPU 包含数量较少但功能强大的内核，具有独立的计算、控制和缓存组件。相反，GPU 拥有许多功率较低的核心，每个核心都包含多个共享公共缓存和控制元素的算术逻辑单元 (ALU)。在图像处理中，拥有多个核心变得至关重要，因为每个像素都需要读取、处理和显示独特的数据。考虑到标准的 1,920x1,080 像素显示屏（相当于 2,073,600 像素），GPU 采用的单指令多数据 (SIMD) 处理变得至关重要，因为它们可以同时执行多个此类操作。与早期模型相比，这提高了计算机游戏的视觉质量。除了图形渲染之外，研究人员还使用 GPU 进行复杂的科学研究，包括基因组测序和分子建模。加密货币挖矿严重依赖 GPU 进行交易验证，而人工智能（尤其是深度学习）则因大型数据集需求而受益匪浅。最新的超级计算机 Frontier 结合了 8,699,904 个 GPU 和 CPU 内核，达到每秒 1.194 exa-flops 的惊人速度。自 1945 年以来，传统 CPU 的发展基本保持不变，GPU 等创新架构的出现使我们能够应对最艰巨的 HPC 挑战。探索 Hive Digital 的环保 GPU 如何满足全球多样化的高性能计算需求。

原文

Graphics Processing Units, or GPUs, have moved beyond their original role of rendering video game graphics and are now used in a variety of high-performance computing applications (HPC), from AI training to zooplankton classification.

To help understand this pivot, Visual Capitalist teamed up with HIVE Digital to look at how GPUs differ from traditional CPUs and what gives them an edge.

CPU vs. GPUs

CPUs, or Central Processing Units, and GPUs, generally have three main elements:

compute elements—technically ALU or arithmetic logic units—that perform calculations and carry out operations;
a control element that coordinates the operations of the above; and
various levels of memory, including dynamic random access memory (DRAM), a kind of RAM or short-term memory used in the main memory of computers, and caches.

CPUs, or Central Processing Units, typically have one or more extremely powerful cores, made up of independent compute, control, and cache elements. A GPU, on the other hand, has many more less-powerful cores, each with multiple ALUs that share common cache and control elements.

Core Values and the Value of Cores

The number of cores is important, especially when it comes to image processing. In order to display an image on your screen, the computer has to read, process, and display data for each pixel, which on modern high-definition displays can really add up. A 1,920 by 1,080 pixel display, for example, has 2,073,600 pixels.

Unlike CPUs, which have to go one operation at a time, GPUs can handle multiple operations like this in parallel, thanks to its multiple-core architecture. Computer scientists call this method of data-handling single instruction, multiple data (SIMD), but all we need to know is that this is why today’s computer games look so much better than 1972’s Pong.

Beyond Graphics

It turns out that GPUs can do more than just render graphics. Researchers are now using GPUs to model protein folding and sequence genomes, while cryptocurrency miners rely on them to validate transactions. GPUs are also playing a critical role in the field of AI, where training datasets are only getting larger.

GPUs are also working side-by-side with CPUs in the world’s only exa-scale computer, Frontier, which uses a combined 8,699,904 GPU and CPU cores to achieve an impressive speed of 1.194 exa-flops per second.

But when you consider that CPUs are still built on roughly the same von Neumann architecture from 1945, it’s perhaps no surprise that new specialized designs, like the GPU, are emerging to help us tackle some of the world’s toughest HPC challenges.

Learn how Hive Digital’s renewably powered GPUs are helping customers worldwide meet their high-performance computing needs.