您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：超算产业需要更多算力中立供应商
记者：郭思
编辑：包永刚
深夜十一点，某新能源汽车企业的仿真计算中心里，屏幕上的流体模拟进度条正在缓缓推进。窗外的城市已经安静下来，但这里的服务器仍在轰鸣——工程师白天做分析、调参数，到了晚上就把计算任务交给算力，第二天一早推开门，结果已经摆在了桌上。
这种“夕发朝看”的工作节奏，正在中国各地的研发实验室里成为新常态。
而与此同时，在内蒙古、宁夏、新疆的广袤土地上，一排排服务器机房亮着指示灯，像沉默的巨人吞吐着数据。这些地方的电费便宜、气候适宜，是部署大规模算力中心的理想之选。但它们不再是过去那种只服务于高校科研的“象牙塔”——越来越多的汽车制造商、大模型企业、航空航天企业、生物医药公司，正在成为这些机房里最活跃的用户。
超算，这个曾经带着浓厚学术色彩的词汇，正在经历一场深刻的产业迁徙。
超算“出圈”：从象牙塔到产业主战场
“五六年前，科研跟高校的场景在我们业务里要超过70%-80%。但到今年，需求更多的是来自企业。”并行科技行业事业部总经理吕昇亮在采访中这样描述行业的巨变。
这个数字背后的驱动力，是各行各业对算力需求的急剧增长。IDC发布的《2026年中国人工智能计算力发展评估报告》预计，2026年中国智能算力规模将达2576.5EFLOPS，同比增长87.9%，到2030年将突破万EFLOPS大关。
面对这一问题，过去，企业要么自建算力集群，要么依托高校超算中心“搭便车”。但这条路如今越走越窄。自建集群投入高、运维重，且难以应对研发高峰期的弹性需求；而高校超算中心本身资源有限，面对企业级任务的体量和时效要求，常常力不从心，曾经就有车企CAE工程师遭遇仿真任务积压两周、排队12天起步的窘境。
“以前我们给高校服务，用户群体比较固定，业务增长相对缓慢。但现在不一样了，”吕昇亮说，“企业端的算力需求在急剧增长，对于需求满足的时效性也在急剧提高”。
这也就意味算力使用模式的变革才是这场迁徙的核心。
过去，企业获取算力的方式很简单——买整机、租整机。但这种方式面临着典型的资源错配困境：买多了浪费，买少了不够用。因为研发工作天然存在潮汐效应——白天工程师们需要即时算力做分析和调试，晚上则可以放下工作让算力继续跑计算。
“我们给客户提了一个叫「夕发朝看」的业务模式，吕昇亮介绍道，“工程师白天在公司看前一天计算的结果，做分
析调整。下班后希望算力不要休息，让算力帮他把计算继续跑下去。第二天早上九点上班的时候，就能看到计算结果。”
为了实现这个模式，并行科技做了一系列复杂的后端工作：保证计算的准确性、出错时自动重跑、智能调度后端资源……“对一个工程师来讲，不用再担心因为算力不够要排队，或者算力多了浪费的问题。你想用多少就有多少，不用的时候也不收费。”
这种按需供给的模式，被吕昇亮比喻为”像用电一样”——开灯就有电，关灯就不计费。
而更深层的变化来自Token经济的兴起。随着大模型和AI应用的爆发，算力的使用模式开始从训练为主转向推理为主。“随着大模型Token调用量快速增长，推理正在成为新增算力的重要驱动力，”吕昇亮说，“我们联合合作伙伴推出了面向推理场景的Token工厂，为不同行业客户的模型推理和智能应用提供算力与平台服务，同时不介入客户模型和应用本身。”
推理场景对算力的要求与训练截然不同，有些场景对实时性要求极高（比如观看短视频、玩在线游戏），需要算力部署在离用户近的地方；而有些场景（比如AI辅助编程）则可以容忍一定的延迟，适合部署在算力成本更低的西部数据中心。
“如果把算力部署在内蒙古、宁夏、新疆，电费比较便宜，整体算力成本就降低了。”吕昇亮分析道，“但如果是实时性要求高的场景，我们就需要把算力部署到离客户近的地方，比如上海、广州。”
从”算得更快”到”算得更好更省”，这不仅仅是技术路线的调整，更是整个算力产业商业模式的深刻重构。
算力黑土地的“超级调度”优势
在这场变革中，并行科技提出并实践的“并行算网”模式，提供了一个独特的解题思路。
截至目前，并行科技已经接入了大概63个算力中心，其中有四十多个智算中心、十几个超算中心。这些分布在全国各地的算力资源，通过一套智能调度系统被整合成一个统一的”算力池”。
“客户提交一个任务上来，我怎么能判断出来他究竟后端的算力哪一个最适合他？”面对这个核心问题。并行科技的答案是一套自主研发的”应用运行特征分析”软件。
并行科技提出一个理念，希望每个新客户来，四个小时之内就要判断出来他究竟是什么样的应用，是计算密集型还是访存密集型，然后根据实际数据和我们的经验，找到性价比最优的计算资源。”
但这套系统面临的挑战远不止技术层面。62个算力中心，有的是自建数据中心，有的是与地方政府合作共建，有的是第三方算力资源。“大家怎么实现一个双赢的模式？要让大家都能挣到钱。”吕昇亮说，“这中间涉及到定价、结算，需要一套技术和商务相结合的模式架构。”
并行科技的策略是做好”黑土地”——只做IaaS和PaaS层的基础架构和平台层，不碰客户的应用层。“我们的客户是做应用的、做大模型研发的、做汽车研发的工程师，他们对自己的业务更了解。我们先把基础架构做好，让他们在上面长出自己的一根烟囱来。”
这种中立第三方的定位，在AI时代显得尤为珍贵。“很多做大模型的厂商，如果把算力完全构建在某个大厂上面，而那个大厂又有自己的大模型，那万一因为竞争原因卡你算力，你的业务就受很大影响。”吕昇亮指出，“所以并行科技更愿意选择一个不做模型、只做算力的中立供应商。”
超1000个核心同时工作，智能汽车折射的AI算力未来
要理解超算如何深入产业，新能源汽车行业是一个绝佳的观察窗口。
“汽车制造是我们事业部营收最大的一块，接近40%的营收来自汽车制造和新能源。”吕昇亮介绍道。一家新能源汽车企业的研发，涉及数千名工程师，从整车设计到自动驾驶算法训练，算力需求贯穿了整个研发链条。
在传统的CAE仿真领域，一台新车的设计需要经过流体仿真（空气动力学）、固体仿真（结构强度）、电池仿真（电化学性能）、整车仿真等多个环节。这些计算需要大规模的并行计算能力，有时需要512个甚至上千个核心同时工作。
“不可能每次造出新车以后去做各种实验，做碰撞、做流体，成本太高了。”吕昇亮说，“所以用计算机辅助工程，用高性能计算来辅助工程设计，这个方法已经发展了五十年。”
但AI的到来给汽车行业带来了全新的算力需求。自动驾驶大模型的训练需要GPU算力，AI辅助编程正在改变软件工程师的工作方式“现在工程师如果还不会用AI来辅助写代码，基本上很难有生存空间。”吕昇亮观察到，“原来一行代码一行代码敲，现在可能是用自然语言告诉AI要实现什么功能，让AI帮忙写代码。”
更前沿的变化来自具身智能。“做智驾的人去做具身，因为他们有经验可以迁移。”吕昇亮分析道，“但具身智能需要的不只是制造业的仿真能力，还需要有‘大脑’和‘小脑’——世界模型、AI场景。所以我们最近在提出把超算跟智算再融合在一起，我们叫超智融合。”
从流体仿真到自动驾驶训练，从AI辅助编程到具身智能，新能源汽车行业对算力的需求正在呈现多元化、指数级增长的态势。而并行算网的价值，就在于能够为这些截然不同的场景，找到最合适的算力配置方案。
而当我们站在产业变革的节点上回望，超算从实验室走向产业一线的趋势已经不可逆转。
“AI能够带来人们实际生产和生活效率的提升，这是毫无疑问的，”吕昇亮说，“目前所存在的场景，这些所谓的AI应用还处在非常初级的阶段。未来三年、五年甚至十年之后，AI催生的工作效率提升可能是我们现在想象不到的。”
当算力从一种稀缺的科研资源，变成像水电一样按需取用的产业基础设施，一场关于生产效率的深层革命正在发生。从”算得更快”到”算得更好更省”，从单一集群到分布式算网，从服务高校到赋能千行百业——超算的产业迁徙，才刚刚开始。
雷峰网(公众号：雷峰网) 雷峰网
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: The overcalculable industry needs more computing and neutral supply Business
Reporter: Gus.
Editor: Pao Yonggang
At 11 p.m., in the simulation centre of a new energy car company, the fluid simulation strip on the screen is slowly advancing. The city outside the window has calmed down, but the servers are still ringing — engineers do analysis and adjust parameters during the day, handing over computing tasks at night and pushing the door the next morning, and the results are already on the table.
This working tempo of “the dawn of dawn” is becoming a new normal in research and development laboratories across China.
At the same time, a row of server rooms in the vast lands of Inner Mongolia, Ningxia, Xinjiang have a signal light on, like silent giants swallowing data. The low cost of electricity and the climate are ideal options for deploying large-scale computing centres. But they are no longer the type of “ivory towers” that used to serve university research only — the growing number of car manufacturers, LLM enterprises, aerospace companies, biomedicine companies — are becoming the most active users in these rooms.
It's super-calculative, a term that used to be very academic, and it's going through a profound industrial migration.
Super circle: from ivory towers to industrial battlefields
Five or six years ago, the scene of scientific research and higher education was over 70-80 percent in our business. But by this year, more needs come from businesses.” The Director-General of the Parallel Science and Technology Industry Department, Rui, highlighted the dramatic changes in the industry as described in the interview.
The driving force behind this figure is the sharp increase in the demand for computing power from all sectors. The IDCpublish Report on the Assessment of China ' s Artificial Intelligence Calculus Development 2026 predicts that China ' s intelligent calculus size will reach 2576.5 EFLOPS in 2026, an increase of 87.9 per cent over the same period, and that by 2030 it will break the 10,000 EFLOPS mark.
Faced with this problem, in the past, enterprises had either built their own clusters of calculators or relied on the Higher Education Supercalculator to “feed off”. But this road is getting narrower. Self-constructing clusters are high, well-functioning and difficult to cope with the need for elasticity during the peak period of R&D; and higher education over-calculation centres have their own limited resources, are often ill-equipped in terms of the volume and timeliness of enterprise-level tasks, and have experienced a two-week backlog and 12-day queue of car company CAE engineers.
“We used to provide tertiary education services with a fixed user base and relatively slow business growth. But it is not the same now,” Lu put it so clearly that “the business needs for computing are growing sharply and the timeliness of meeting demand is increasing dramatically”.
This also means that changes in the mode of computing use are at the heart of this migration.
In the past, the way in which businesses had gained their ability to calculate was simple — to buy the whole machine, to rent it. This approach, however, faces typical resource mismatches: more is wasted and less is not enough. Because R&D has a natural tidal effect — the day engineers need instant calculus for analysis and debugging, and at night they can put down their jobs to keep them running.
"We gave the client a business model called "The Sunset Eyes", "The engineers read the results of the calculations the day before at the company and scored them."
Parsing adjustment. After work, I hope I don't rest so I can help him keep running. At 9:00 a.m. the next morning, you can see the calculations."
In order to achieve this model, parallel technologies do a complex set of back-end work: ensuring accuracy of calculations, automatic re-runs in case of error, intelligent movement of back-end resources... “For an engineer, there is no need to worry about queues because of insufficient or wasteful calculations. You can use as much as you want, and you don't charge for it."
This supply-on-demand pattern has been highlighted as “like electricity” — lights have electricity and lights are free of charge.
The deeper changes came from the rise of the Token economy. With the outbreak of LLM and AI applications, the mode of use of algorithms began to shift from training to reasoning. “As LLMToken's call grew rapidly, reasoning is becoming an important driver of the new calculus.” Lu put it out, “Our joint partners launched the Token factory, which is oriented towards the reasoning scene, providing algorithms and platform services for model reasoning and smart applications for different industry clients, without intervening in client models and applications themselves.”
Delineation scenarios require a very different level of computing than training, some of which are extremely demanding in real time (e.g. watching short video, playing online games) and need to be deployed in close proximity to users; others (e.g. AI-assisted programming) can tolerate a certain delay and are suitable for deployment in a lower-cost Western data centre.
“If power is deployed in Mongolia, Ningxia, Xinjiang, electricity is cheaper, the overall cost of computing is lower.” Lu's analysis is bright, “but if real-time is a high-demand scenario, we need to deploy it to places close to customers, such as Shanghai and Guangzhou.”
From “calculating faster” to “calculating better and less”, this is not just a re-routing of technology, but a profound re-engineering of the business model of the entire computing industry.
"Super dispatch" advantage for dark land
In this transformation, parallel technologies have introduced and put into practice a “parallel” model that provides a unique solution.
To date, parallel technologies have been connected to approximately 63 computing centres, including more than 40 think centres and more than a dozen supercentres. These computing resources, distributed throughout the country, are integrated into a unified “calculator pool” through a smart dispatch system.
"How can I determine which of his back-end calculations is best for him?" Faced with this central issue. The answer to parallel technology is an autonomous set of applied operational profiling software.
Parallel technology offers the idea that every new customer should come in within four hours to determine what kind of application he is, whether he is computing or visiting, and then find the best cost-effective computing resources based on actual data and our experience.”
However, the challenges facing the system are far more than technical. Sixty-two computing centres, either self-established data centres or joint projects with local governments or third-party computing resources. “How can we achieve a win-win model? We need to make money for everyone." “This involves pricing, settlement and requires a model structure that combines technology and commerce.”
The strategy of parallel technology is to do "black land" – only basic structures and platforms at the Iaas and PaaS levels, without touching the client’s application level. “Our clients are engineers who do applications, do LLM, do auto research and development, and they know their business better. Let's start with the basics and let them grow their own chimneys."
This location of a neutral third party is particularly valuable in the AI era. “Many producers who do LLM, if they have their own LLM fully built up on a large plant, will be greatly affected by your business if it's too competitive for you to do so.” Lu Lu Xinjiang pointed out, “So parallel technology prefers to choose a neutral supplier that does not model, but only counts.”
Over 1,000 cores working at the same time.
To understand how supercalculations go deep into industry, the new energy automobile industry is an excellent observation window.
“Vehicle manufacturing is the largest part of our career, almost 40 per cent of which comes from auto manufacturing and new energy.” Lu's introduction is bright. The research and development of a new energy automobile enterprise, involving thousands of engineers, spanned the entire research and development chain from the design of a car to the training of autopilot algorithms.
In the traditional CAE simulation area, a new vehicle needs to be designed using fluid simulation (aerodynamics), solid simulation (structure strength), battery simulation (electrical chemistry performance) and whole vehicle simulation. These calculations require large parallel computing capabilities, sometimes 512 or even thousands of cores working simultaneously.
“It is impossible to do experiments, collisions, fluids every time a new car is built. Lu stated, “So computer-aided engineering, supported by high performance calculations, has been developed for 50 years.”
But the arrival of AI has brought new computing demands to the automobile industry. The training for auto-piloting LLM requires GPU calculus, and AI-assisted programming is changing the way software engineers do their work: “It is now very difficult for engineers to have room for survival if they do not use AI-aided code.” Lu has observed, “The original line of code knocks, and now it may be in natural languages to tell AI what to do and allow AI to write codes.”
The more advanced changes come from intelligence. “A person who is smart enough to do a body, because they have experience in moving.” Lu's analysis says, “But what is smart enough to do is not just a simulation of manufacturing, but also a `brain' and `brain' — a world model, an AI scene. So we've recently offered to merge super and smart. We're called super.
From fluid simulation to self-driving training, from AI-aided programming to smartness, the demand for computing power in the new energy automobile industry is diversifying and increasing exponentially. The value of a parallel network lies in the ability to find the most appropriate calculator configuration for these distinct scenarios.
And when we look back at the nodes of industrial change, the trend of moving beyond laboratory to industry is irreversible.
There is no doubt that AI can bring about an increase in people's productivity and efficiency in real life, and Lu states that “the scenes that exist are at a very early stage of so-called AI applications. In the next three years, five years or even 10 years, the efficiency gains that AI has generated may not be within our grasp.”
An in-depth revolution on production efficiency is taking place as the force moves from a scarce scientific resource to a needs-based industrial infrastructure like hydropower. From "calculated faster" to "calculated faster" to "calculated better" and less costly, from a single cluster to a distributed computing network, from a service college to an enabling business — an overcalculable industrial migration — has only just begun.
Thunder Peak Network.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.