谷歌开源 TorchTPU 挑战英伟达 AI 生态主导地位Google Open-Sources TorchTPU to Challenge Nvidia's AI Ecosystem Dominance


AI free token/AI免费积分

IT之家 10 月 21 日消息,谷歌于当地时间周二宣布,面向谷歌 TPU 的原生 PyTorch 框架 TorchTPU 现已在 GitHub 上开源。借此,谷歌与亚马逊一同尝试打破英伟达在人工智能行业的垄断。当下绝大多数人工智能应用默认都是基于英伟达芯片,也就是图形处理器(GPU)开发运行。

谷歌开源 TorchTPU 挑战英伟达 AI 生态主导地位

PyTorch 是一款热门开源工具库,人工智能应用开发者与初创企业依靠它搭建、训练和部署人工智能模型。有统计称,85% 的人工智能应用都在使用 PyTorch。PyTorch 最初由 Meta 在 2016 年开发,那时人们还完全没有意识到大语言模型将会改变世界。当时英伟达 GPU 已经是人工智能研究的标准硬件,因而 PyTorch 的诸多能力都是专门为适配英伟达 GPU 构建。2022 年,Meta 将 PyTorch 移交至 Linux 基金会,这家非营利机构负责管理许多关键的开源项目。

谷歌推出的 TorchTPU,可以让人工智能应用在谷歌自研人工智能芯片张量处理单元(TPU)上运行。它还可以便捷地将现有的 PyTorch 应用迁移到 TPU,进而迁移到谷歌云。开发者仅需简单修改规格,无需重写应用代码即可完成迁移。

亚马逊在当地时间周一发布了类似消息,推出自家的 PyTorch 版本 TorchNeuron。TorchNeuron 可以在无需重写代码的情况下,把人工智能应用移植到亚马逊自研芯片 Trainium 以及亚马逊云(AWS)。亚马逊还表示,将进一步削弱英伟达对 PyTorch 的掌控力。其正在向 PyTorch 主项目提交代码,把各类能力与英伟达芯片解耦,让 PyTorch 原生就能兼容各类硬件。

不过IT之家注意到,谷歌在部分维度的举措力度或许超过亚马逊。谷歌称,其直接联合 Meta、Hugging Face、英菲拉克特(Inferact,vLLM 开发方)、拉迪克斯阿克(RadixArk,SGLang 开发方)等多方合作,保证 TorchTPU 的能力与代码库可以像原版 PyTorch 一样,和这些主流人工智能开发工具良好兼容。

除此之外,这家搜索巨头推出了专用工具 TPU 智能体套件,协助开发者把应用迁移到谷歌芯片与云平台,并在迁移过程中排查修复错误。谷歌还升级了广受使用的 LiteRT 工具,该工具用于在手机、笔记本电脑和嵌入式设备上运行人工智能模型,升级后可以更好地兼容基于 PyTorch 开发的人工智能应用。

以上这些进展,并不代表人工智能应用开发者会仅仅因为可以便捷迁移到其他芯片与云平台,就集体放弃英伟达硬件。但亚马逊和谷歌显然正在努力终结英伟达的长处地位。

本文来源:https://www.ithome.com/1/015/626.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT Home October 21st News, Google announced on Tuesday, local time, that TorchTPU, the native PyTorch framework for Google TPU, is now open source on GitHub. In doing so, Google joined Amazon in an attempt to break Nvidia's monopoly in the AI industry. The vast majority of AI applications today are developed and run on Nvidia chips, or graphics processors (GPUs), by default.

PyTorch is a popular open source toolkit that AI app developers and startups rely on to build, train, and deploy AI models. According to some statistics, 85% of AI applications are using PyTorch. Originally developed by Meta in 2016, PyTorch was completely unaware that big language models would change the world. At the time, Nvidia GPUs were already the standard hardware for AI research, so many of PyTorch's capabilities were built specifically for Nvidia GPUs. In 2022, Meta handed over PyTorch to the Linux Foundation, a nonprofit that manages many key open source projects.

Google's TorchTPU allows AI applications to run on Google's self-developed artificial intelligence chip tensor processing unit (TPU). It also makes it easy to migrate existing PyTorch applications to TPUs and then to Google Cloud. Developers can simply modify the specifications and complete the migration without rewriting the application code.

Amazon released a similar message on Monday with its own PyTorch version of TorchNeuron. TorchNeuron can port AI applications to Amazon's self-developed chip Trainium and Amazon Cloud (AWS) without having to rewrite the code. Amazon also said it would further weaken Nvidia's hold on PyTorch. It is submitting code to the PyTorch main project to decouple various capabilities from Nvidia chips, making PyTorch natively compatible with various types of hardware.

However, IT House noticed that Google's efforts in some dimensions may be stronger than Amazon's. Google said that its direct cooperation with Meta, Hugging Face, Inferact (vLLM developer), RadixArk (SGLang developer) and other parties ensures that TorchTPU's capabilities and codebase can be compatible with these mainstream AI development tools like the original PyTorch.

In addition, the search giant launched a suite of specialized TPU agents to help developers migrate applications to Google's chip and cloud platforms and troubleshoot bugs during the migration process. Google has also upgraded its widely used LiteRT tool, which runs AI models on phones, laptops and embedded devices, to be more compatible with PyTorch-based AI applications.

These developments do not mean that AI application developers will collectively abandon Nvidia hardware simply because it can be easily migrated to other chips and cloud platforms. But Amazon and Google are clearly working to end Nvidia's strengths.

This article is from: https://www.ithome.com/1/015/626.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

Google Open-Sources TorchTPU to Challenge Nvidia's AI Ecosystem Dominance

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry