【观点】高文院士：发展算力网要迎接直连和调度两大挑战

8月19日举行的2023中国算力大会、第二届西部数谷算力产业大会主论坛上，中国工程院院士、50人论坛委员高文以《智能算力助力数字经济高质量发展》为题发表演讲。

高文认为，算力不仅对数字经济很重要，其实对GDP的发展就非常重要。此前清华大学和IDC咨询联合发布的《2022-2023全球计算力指数评估报告》显示，2021年、2022年全世界GDP和算力是正相关的。“美国计算力指数比中国多，中国大概是美国的70%左右，而GDP和美国也差这样的距离，排在后面的日本等这些国家GDP排序完全和算力指数排序是对等的。换句话说，算力就是生产力，有算力就会有GDP。总而言之，如果是算力不行，GDP肯定不行。怎么样把算力转换成GDP，这里有一定的文章可做。”

中国在算力增长方面是全世界最强劲的。高文表示：“我们的算力增长的速度是年均13.5%，而美国只有5%，其他国家都没有我们强。”高文判断，未来算力的强力供给，很可能会像今天的电力一样，随着算力到处输出，使得整个经济发展速度非常快。“当然，为了做到这点需要有算力网络，只有算力还不行，要把算力输送到最需要的地方。”

高文认为，中国“东数西算”其实有点像“西电东送”，是希望把西面算力送到东面，最后使得这些算力低成本地供给给需要方，这当然是一个非常好的想法。“这件事要想做到，要像提供电网一样提供算力网，然后把算力送过去。所以要直送或者以最经济、最便宜的办法把算力送过去，让需要方用到的和在本地用到的是一样的，这是一个非常好的目标，我们把这样的目标叫作中国算力网。”简单点说，如果可以把算力像电力一样送到需要的地方，这个网就叫算力网。 CXO UNION-CXO联盟（cxounion.cn）

好的算力网需要迎接诸多技术挑战。高文认为，第一个挑战是连接挑战，具体涉及连接延迟和超宽带宽问题。就连接延迟而言，算力网是时间敏感网络，“比如将宁夏银川、中卫的算力送到福建、青岛，最小延迟是多少？网络连接的最小物理延迟现在是每200公里需要1毫秒，这是在用光纤直连、中间没有任何交换的情况下，实际上这个延迟在中途可能会有些放大。每200公里1毫秒，2000公里就是10毫秒，4000公里就是20毫秒，这是现在物理上能做到的。”高文认为，延迟优化还有提升空间。即把光纤变成空心光纤，就能实现300公里1毫秒，这也是实验室里正在攻关的课题。就超宽带宽来看，要想把数据非常通畅地送来送去，带宽要够。而鹏城国家实验室一个主攻课题就是要提供100T-1000T的带宽技术。 CXO UNION-CXO联盟（cxounion.cn）

第二个挑战是调度。算力能不能调度，能调度到什么程度？至少现阶段调度对于“西电东送”是很重要的，包括绿电能不能上网。对于算力来讲，很重要的一点是和调度关联的能力能不能相位对上，或者能不能在需要的时候上，不需要的时候别上，“不能不需要的时候上了很多，需要的时候没有，所以调度问题很关键。”高文认为，算力的调度尤其难。

据高文介绍，迄今算力有三类。一是云算力，比如说现在宁夏中卫、银川做的很多算力项目，是运营商或者大的云服务商提供的云算力。这类算力很重要。第二类叫智能算力，这是最近几年刚出来的，比如训练大模型需要的智能算力。第三类叫超级算力或超算算力，主要涉及国内很多超算中心的算力。“这三类不同算力配送的时候完全是不同体系架构，因为用户类型是不一样，所以现在只能分类输送，也就是说分成三种不同的算力封装输送。”高文认为，三种不同的算力要怎样进行调度是现阶段面临的挑战。远期来看，他判断希望在于混合算力，“火电、太阳能发电、风电同样都从电网送过去，虽然现阶段的算力得分类输送，但将来也是要混合输送的。”

要想做好混合输送，高文强调，第一点是全国大的算力中心之间一定要有超宽带宽、超低延迟的网络直连。“这叫算力信息高速公路或者叫算力高速公路。连接靠高速公路，而算力调度靠云原生网络。云原生是近几年很重要的概念，有云原生软件、云原生网络，这是算力调度起来最关键的核心技术，我们一定要进入云原生网络时代。”

云原生网络最核心的是云。以前最核心的是网络，以后会是云。因为现在的网络是由三个网络构建起来的，比如IP网、接入网、主干网。以后云原生把三个网融合到一起，其中最核心的就是云。所以要把IP承载网、无线接入网这些都“融化”到云里去，数据中心的网也是。 CXO UNION-CXO联盟（cxounion.cn）

第二点是算力的进步。高文介绍到，目前鹏城国家实验室正在提升算力供给能力，已经完成了云脑2项目，拥有1000P算力，相当于50万台服务器。“我们现在做云脑3，具有16000P算力，相当于800万台服务器，这个算力将来会提供给‘东数西算’国家智能超算使用，为了做这个事我们设计了中国算力网。”据高文介绍，中国算力网已经迭代了两个版本：先C2NET 0.1做了很多标准化的工作，也做了长距离的连接。长距连接已经完成的110T，可以传输2000公里的实验在实验室里做完了；现在正在做的是C2NET 1.0，希望能把国内“东数西算”工程节点以及其他智能节点连接起来，用“特高压”的概念把所有的算力节点连起来提供使用，可以跨地域，也可以做异构的一些使用。

高文总结到，要让真正的算力、中国的算力能够发挥它的最大效用，对GDP直接产生作用，就要解决两个技术难题：一是直连，包括高速直连和超宽带宽直连；二是算力调度。要解决好这些问题，需要各方携手共进。

翻译：

Academician Gao Wen: The development of computing power network should meet the two major challenges of direct connection and scheduling

At the main forum of the 2023 China Computing Power Conference and the second Western Digital Valley Computing Industry Conference held on August 19, Gao Wen, academician of the Chinese Academy of Engineering and member of the 50-member forum, delivered a speech on the topic of “Intelligent Computing Power Helps the high-quality development of the digital Economy”.

Gowen believes that computing power is not only important for the digital economy, but also for the development of GDP. The “2022-2023 Global Computing Power Index Evaluation Report” jointly released by Tsinghua University and IDC Consulting shows that the world’s GDP and computing power are positively correlated in 2021 and 2022. “The United States computing power index is more than China, China is about 70% of the United States, and GDP and the United States are also such a distance, ranked behind Japan and other countries GDP ranking is completely equivalent to the computing power index ranking.” In other words, computing power is productivity, and with computing power comes GDP. In short, if computing power is not good, GDP is certainly not good. There are certain articles on how to convert computing power into GDP.”

China is the world’s strongest in terms of computing power growth. “Our computing power is growing at 13.5 per cent a year, compared to 5 per cent in the US and no other country,” Mr Gowen said. Gowen reckons that the future’s powerful supply of computing power is likely to be like electricity today, as computing power is exported everywhere, making the entire economy grow very fast. “Of course, in order to do this, you need to have a network of computing power, only computing power is not enough, to deliver computing power to where it is most needed.” CXO UNION-CXO联盟（cxounion.cn）

Gao Wen believes that China’s “east to West calculation” is actually a bit like “west to east power transmission”, is hoping to send the western computing power to the east, and finally make these computing power at low cost supply to the need, which is of course a very good idea. “If you want to do this, you have to provide a computing power network like a power grid, and then send the computing power over.” So to send directly or in the most economical and cheapest way to send the computing power, so that the need and the local use is the same, this is a very good goal, we call such a goal called China Computing power network.” To put it simply, if the computing power can be sent to the place where it is needed like electricity, the network is called the computing power network.

A good computing network needs to meet many technical challenges. According to Gowen, the first challenge is the connectivity challenge, which relates specifically to connection latency and ultra-wide bandwidth issues. In terms of connection delay, the computing power network is a time-sensitive network, “For example, the computing power of Yinchuan and Zhongwei in Ningxia is sent to Fujian and Qingdao, what is the minimum delay?” The minimum physical latency for a network connection is now 1 millisecond per 200km, and that’s when there’s no exchange between direct fiber connections, which can actually get a little bigger in the middle. Every 200 kilometers is a millisecond, 2,000 kilometers is 10 milliseconds, 4,000 kilometers is 20 milliseconds, and that is what is physically possible today.” Gowen believes there is room for improvement in delay optimization. That is, turning the fiber into a hollow fiber can achieve 300 kilometers for 1 millisecond, which is also a topic that is being tackled in the laboratory. In terms of ultra-broadband width, in order to send data to and from very unobstructed, the bandwidth must be enough. And Pengcheng National Laboratory a main topic is to provide 100T-1000T bandwidth technology.

The second challenge is scheduling. Can the computing power be scheduled, and to what extent can it be scheduled? At least at this stage, dispatch is very important for “west-east power transmission”, including green power can not be connected to the Internet. For computing power, a very important point is whether the ability associated with scheduling can be phase-matched, or whether it can be on when it is needed, and don’t be on when it is not needed, “There is a lot of it when it is needed, and there is no it when it is needed, so the scheduling problem is very critical.” According to Gowen, the scheduling of computing power is particularly difficult. CXO UNION-CXO联盟（cxounion.cn）

So far, according to Gawain, there are three types of computing power. First, cloud computing power, for example, many computing power projects done in Ningxia Zhongwei and Yinchuan are cloud computing power provided by operators or large cloud service providers. This kind of computing power is important. The second category is called intelligent computing power, which has just come out in recent years, such as the intelligent computing power needed to train large models. The third category is called super computing power or supercomputing power, which mainly involves the computing power of many domestic supercomputing centers. “These three different computing power distribution is completely different architecture, because the type of user is not the same, so now can only be classified transportation, that is, divided into three different computing power packaging transportation.” Gowen believes that how to schedule three different computing forces is the challenge at this stage. In the long term, he judged that the hope lies in the hybrid computing power, “thermal power, solar power, wind power are also sent from the grid, although the current computing power has to be classified, but in the future it will also be mixed.”

In order to do a good job of mixed transmission, Gao Wen stressed that the first point is that there must be ultra-wide broadband and ultra-low latency network direct connection between the country’s large computing power centers. “It’s called the computing information Superhighway or the computing superhighway. Connectivity is on the highway, and computing power scheduling is on the cloud-native network. Cloud native is a very important concept in recent years, there are cloud native software, cloud native network, which is the most critical core technology of computing power scheduling, we must enter the cloud native network era.” CXO UNION-CXO联盟（cxounion.cn）

The core of cloud-native networks is the cloud. In the past, the core was the network, and in the future it will be the cloud. Because the current network is built by three networks, such as IP network, access network, backbone network. In the future, the cloud native integrates the three networks together, the most core of which is the cloud. Therefore, it is necessary to “melt” the IP bearer network and wireless access network to the cloud, and the data center network is also.

The second point is the advancement of computing power. Gao Wen introduced that at present, Pengcheng National Laboratory is improving its computing power supply capacity, and has completed the Cloud brain 2 project, which has 1000P computing power, equivalent to 500,000 servers. “We are now doing cloud brain 3, with 16000P computing power, equivalent to 8 million servers, this computing power will be provided to the ‘East number west calculation’ national intelligent supercomputer use, in order to do this we designed the Chinese computing power network.” According to Gao Wen, China’s computer power network has iterated two versions: first C2NET 0.1 did a lot of standardization work, but also did long-distance connections. Long distance connection has been completed 110T, can be transmitted 2000 kilometers of the experiment done in the laboratory; What is being done now is C2NET 1.0, hoping to connect the domestic “east number west calculation” project nodes and other intelligent nodes, and use the concept of “UHV” to connect all the computing power nodes to provide use, which can be cross-regional and can also be used heterogeneous. CXO UNION-CXO联盟（cxounion.cn）

Gao Wen concluded that in order for real computing power, China’s computing power, to play its maximum utility and have a direct effect on GDP, two technical problems must be solved: first, direct connection, including high-speed direct connection and ultra-wide broadband direct connection; The second is computing power scheduling. To solve these problems, all parties need to work together.

由CXO UNION-CXO联盟（cxounion.cn）转载而成，来源于人民邮电报；编辑/翻译：CXO UNIONCXO联盟小U。

如需加入CXO UNION（CXO联盟）高管社群，请联系社群小伙伴哦~

免责声明: 本网站(http://www.cxounion.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。

本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。