我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
算力版隆中对
视频
音频
原始脚本
算力版隆中对，拆解美国创世纪计划的虚实，看清中美算力攻坚的核心棋局。 世人皆知曹操挥师83万下江南，旌旗蔽日，引得孙刘正恐。 唯有诸葛亮一眼看穿虚实，所谓百万大军，不过是青州精锐10余万，收编水军20万，余下皆是充数民夫杂役，看似声势浩大，实则精锐寥寥。 今日美国高调推出的创世纪计划，恰似当年曹军南下。 140PB 海量数据，24家顶尖企业协同，AI 赋能前沿科研的角头震天，引得舆论热议。 但若如孔明般抽丝剥茧拆解，便知其造势大于实效。 而藏在噱头背后的存量超算改造才是真正的硬核杀招。 恰如曹军精锐，值得我们战略上藐视其虚，战术上重视其实，找准自身短板精准破局。 不被声势所惑，亦不被差距所困。 创世纪计划的虚，藏在层层包装的宣传口径里，与曹操83万大军的注水逻辑如出一辙。 其一，140PB数据是账面数字泡沫，如同曹军里的民夫杂役，徒有体量而无核心价值。 这批横跨数十年积累的存量数据，是美国国家实验室过往实验的冗余记录，废弃数据与精度不达标的历史遗存。 绝大多数是无科研价值的数据垃圾，真正能支撑核聚变、量子物理等前沿研究的高价值新数据与精选旧数据不过是其中零头。 所谓全量挖掘，不过是用来造势争资的宣传话术，与当年曹操以民夫充数壮大声势别无二致。 其二，24家企业协同是轻量级占位之举，无强制约束，亦无真金白银重仓。 OpenAI、DeepMind、IBM、Oracle 等企业签下的不过是一纸谅解备忘录，本质是顺势布局的政策公关，开放现有模型，共享云资源，而非新增百亿级专项投入。 核心团队依旧聚焦自身前沿研发，旧数据挖掘不过是实习生级别的 顺手之活。 正如曹军收编的水军，看似人数众多，实则战力参差不齐，难成核心攻坚之力。 其三，政府投入是口惠实不至，财力捉襟见肘，难成大局。 美国政府官宣的首批投入仅3.2亿美元，与此前传闻的8500亿美元总投入相去甚远。 叠加债台高筑，能源部预算遭削减的现实，后续持续拨款毫无保障。 所谓举国攻坚，不过是圣诞假期前放出的政治烟花，热闹过后只剩一地鸡毛。 终究是雷声大，雨点小。 拨开宣传迷雾，创世纪计划的实，藏在存量超算改造的硬核布局里。 这才是美国政府在财力能源双重受限下，为数不多能撬动的关键变量，亦是英伟达等企业破局困局的核心抓手。 美国当下正面临两大无解困境，一是 AI 算力能源瓶颈凸显。 据测算，2026年新增 AI 算力中心电力缺口将达30%~40%，德州、亚利桑那等算力聚集区已现限电预警。 动则年耗电超10亿度的 GPU 集群，已成无法承受的电老虎。 二是英伟达等企业增量市场遇阻，中国市场大门紧闭，叠加美国出口经令限制，全球增量算力需求难以为继，存量市场成了唯一的增长曲线。 在此背景下，改造国家实验室的传统超算成了顺理成章的双赢之选。 橡树岭、劳伦斯利弗莫尔等顶级实验室的超算，本就是国家投巨资打造的算力枢纽。 配套专属核电、水电，能耗成本远低于商业算力中心。 且具备扎实的大规模并行计算底子，此前紧缺 AI 易购算力适配。 如今用英伟达 GB200等芯片做 AI 化升级，本质是给存量优质算力装智能大脑，既无需新建高能耗集群，破解能源困局。 又能快速补齐国家实验室的 AI 算力短板，更能让英伟达从卖增量集群转向吃存量改造，对冲市场下滑压力。 这种政府盘活存量资源、企业挖掘存量市场的绑定，才是创世纪计划真正的硬核内核。 看似低调，实则精准击中了美国当下的痛点，是务实且高效的破局之法。 看清美国计划的虚实之后，便知我们当秉持战略上藐视、战术上重视的核心准则，藐视其宣传造势的虚张声势。 重视其存量超算与 AI 融合的底层逻辑，更要直面自身的短板与困境，走出一条自主可控的算力攻坚之路。 不可否认，中国在算力协同与超算改造上，短板同样突出。 其一，传统超算并行计算体系繁琐落后，堪称科研与产业的拦路虎。 早年从事过 MPI 编程的人都深知，传统超算以 CPU 为核心并行调度依赖复杂的专属编程体系，门槛之高远超普通程序员与科研工作者的能力范畴。 大量顶尖超算的峰值算力最终沦为束之高阁的数字，难以转化为实际科研成果。 其二，超算、AI 企业、国家实验室三者割裂，形成信息孤岛。 我们不缺顶尖超算，神威、天河系列，不缺头部 AI 企业，华为、字节等，也不缺实力雄厚的国家实验室。 但三者长期各自为战，国家实验室有数据、有超算，却缺好用的 AI 模型。 AI 企业有模型、有 AI 算力力，却缺核心科研数据。 据与超算支撑，算力、数据、模型无法形成闭环，难以发挥1+1>2的协同效应。 其三，若盲目效仿美国路径，极易陷入卡脖子困境。 英伟达的芯片垄断，海外架构的技术壁垒，都是悬在头顶的达摩克利斯之剑。 面对美国的布局与自身的短板，中国的应对之策既要精准对标，更要立足自主，不搞一刀切，不随波逐流，聚焦核心痛点稳步推进。 首要之事是攻坚传统超算的并行化改造，破解繁琐难用的核心痛点。 改造的核心绝非简单加装 AI 芯片，而是底层架构的重构与门槛的降低。 以国产 AI 芯片、华为昇腾、海光深算等为核心，搭建 CPU 加国产 AI 芯片的异构并行架构，重构底层调度系统，让高精度科学计算与高吞吐 AI 计算高效协同。 同时搭建低代码并行编程平台，让普通科研人员无需精通 复杂并行算法，就能便捷调用超算与 AI 算力，让顶尖算力真正能用、好用、用出成果。 其次是打通超算加国家实验室加 AI 企业的协同通道，打破信息孤岛。 以政府为统筹，建立数据分级开放机制，将国家实验室非涉密的高价值科研数据合规开放给国产 AI 企业。 打造国家级 AI for Science 协同平台，推动国产科学大模型与超算深度适配。 比如让盘古大模型对接大连化物所、高能所的科研需求，形成数据为模型，模型用算力，算力出成果的正向循环。 让国有存量资源与民营创新活力深度融合。 最后，也是最关键的一点，是坚守自主可控的底线，筑牢底层根基。 从超算架构到 AI 芯片，从并行调度系统到科学大模型，全程立足国产技术，不依赖海外供应链与架构体系，避免走改造及受制的老路，让算力攻坚的每一步都踩在自主可控的土地上，既破解自身短板，又规避卡脖子风险。 当年诸葛亮拆解曹操83万大军，终助孙刘连兵。 赤壁破敌 今日我们拆解美国创世纪计划的虚实，并非为了冷眼旁观，而是为了找准方向，精准发力。 美国的存量超算改造是财力与能源受线下的务实之选，中国的算力攻坚是直面短板、自主创新的必然之路。 无需被美国的宣传声势吓到，亦不能忽视其底层逻辑的可取之处。 战略上保持定力，战术上精准施策，盘活自身存量资源，打通协同壁垒，坚守自主底线，便能让中国的算力优势真正转化为科研优势、产业优势。 在这场全球算力攻坚的棋局中，走出属于我们自己的制胜之路。
修正脚本
算力版隆中对，拆解美国创世纪计划的虚实，看清中美算力攻坚的核心棋局。 世人皆知曹操挥师83万下江南，旌旗蔽日，引得孙刘震恐。 唯有诸葛亮一眼看穿虚实，所谓百万大军，不过是青州精锐10余万，收编水军20万，余下皆是充数民夫杂役，看似声势浩大，实则精锐寥寥。 今日美国高调推出的创世纪计划，恰似当年曹军南下。 140PB 海量数据，24家顶尖企业协同，AI 赋能前沿科研的号角震天，引得舆论热议。 但若如孔明般抽丝剥茧拆解，便知其造势大于实效。 而藏在噱头背后的存量超算改造才是真正的硬核杀招。 恰如曹军精锐，值得我们战略上藐视其虚，战术上重视其实，找准自身短板精准破局。 不被声势所惑，亦不被差距所困。 创世纪计划的虚，藏在层层包装的宣传口径里，与曹操83万大军的注水逻辑如出一辙。 其一，140PB数据是账面数字泡沫，如同曹军里的民夫杂役，徒有体量而无核心价值。 这批横跨数十年积累的存量数据，是美国国家实验室过往实验的冗余记录，废弃数据与精度不达标的历史遗存。 绝大多数是无科研价值的数据垃圾，真正能支撑核聚变、量子物理等前沿研究的高价值新数据与精选旧数据不过是其中零头。 所谓全量挖掘，不过是用来造势争资的宣传话术，与当年曹操以民夫充数壮大声势别无二致。 其二，24家企业协同是轻量级占位之举，无强制约束，亦无真金白银重仓。 OpenAI、DeepMind、IBM、Oracle 等企业签下的不过是一纸谅解备忘录，本质是顺势布局的政策公关，开放现有模型，共享云资源，而非新增百亿级专项投入。 核心团队依旧聚焦自身前沿研发，旧数据挖掘不过是实习生级别的顺手之活。 正如曹军收编的水军，看似人数众多，实则战力参差不齐，难成核心攻坚之力。 其三，政府投入是口惠实不至，财力捉襟见肘，难成大局。 美国政府官宣的首批投入仅3.2亿美元，与此前传闻的8500亿美元总投入相去甚远。 叠加债台高筑，能源部预算遭削减的现实，后续持续拨款毫无保障。 所谓举国攻坚，不过是圣诞假期前放出的政治烟花，热闹过后只剩一地鸡毛。 终究是雷声大，雨点小。 拨开宣传迷雾，创世纪计划的实，藏在存量超算改造的硬核布局里。 这才是美国政府在财力能源双重受限下，为数不多能撬动的关键变量，亦是英伟达等企业突破困局的核心抓手。 美国当下正面临两大无解困境，一是 AI 算力能源瓶颈凸显。 据测算，2026年新增 AI 算力中心电力缺口将达30%~40%，德州、亚利桑那等算力聚集区已现限电预警。 动辄年耗电超10亿度的 GPU 集群，已成无法承受的电老虎。 二是英伟达等企业增量市场遇阻，中国市场大门紧闭，叠加美国出口禁令限制，全球增量算力需求难以为继，存量市场成了唯一的增长曲线。 在此背景下，改造国家实验室的传统超算成了顺理成章的双赢之选。 橡树岭、劳伦斯利弗莫尔等顶级实验室的超算，本就是国家投巨资打造的算力枢纽。 配套专属核电、水电，能耗成本远低于商业算力中心。 且具备扎实的大规模并行计算底子，此前紧缺 AI 异构算力适配。 如今用英伟达 GB200等芯片做 AI 化升级，本质是给存量优质算力装智能大脑，既无需新建高能耗集群，破解能源困局，又能快速补齐国家实验室的 AI 算力短板，更能让英伟达从卖增量集群转向吃存量改造，对冲市场下滑压力。 这种政府盘活存量资源、企业挖掘存量市场的绑定，才是创世纪计划真正的硬核内核。 看似低调，实则精准击中了美国当下的痛点，是务实且高效的破局之法。 看清美国计划的虚实之后，便知我们当秉持战略上藐视、战术上重视的核心准则，藐视其宣传造势的虚张声势。 重视其存量超算与 AI 融合的底层逻辑，更要直面自身的短板与困境，走出一条自主可控的算力攻坚之路。 不可否认，中国在算力协同与超算改造上，短板同样突出。 其一，传统超算并行计算体系繁琐落后，堪称科研与产业的拦路虎。 早年从事过 MPI 编程的人都深知，传统超算以 CPU 为核心并行调度依赖复杂的专属编程体系，门槛之高远超普通程序员与科研工作者的能力范畴。 大量顶尖超算的峰值算力最终沦为束之高阁的数字，难以转化为实际科研成果。 其二，超算、AI 企业、国家实验室三者割裂，形成信息孤岛。 我们不缺顶尖超算，神威、天河系列，不缺头部 AI 企业，华为、字节等，也不缺实力雄厚的国家实验室。 但三者长期各自为战，国家实验室有数据、有超算，却缺好用的 AI 模型。 AI 企业有模型、有 AI 算力，却缺核心科研数据，既无超算支撑，算力、数据、模型无法形成闭环，难以发挥1+1>2的协同效应。 其三，若盲目效仿美国路径，极易陷入卡脖子困境。 英伟达的芯片垄断，海外架构的技术壁垒，都是悬在头顶的达摩克利斯之剑。 面对美国的布局与自身的短板，中国的应对之策既要精准对标，更要立足自主，不搞一刀切，不随波逐流，聚焦核心痛点稳步推进。 首要之事是攻坚传统超算的并行化改造，破解繁琐难用的核心痛点。 改造的核心绝非简单加装 AI 芯片，而是底层架构的重构与门槛的降低。 以国产 AI 芯片、华为昇腾、海光深算等为核心，搭建 CPU 加国产 AI 芯片的异构并行架构，重构底层调度系统，让高精度科学计算与高吞吐 AI 计算高效协同。 同时搭建低代码并行编程平台，让普通科研人员无需精通复杂并行算法，就能便捷调用超算与 AI 算力，让顶尖算力真正能用、好用、用出成果。 其次是打通超算加国家实验室加 AI 企业的协同通道，打破信息孤岛。 以政府为统筹，建立数据分级开放机制，将国家实验室非涉密的高价值科研数据合规开放给国产 AI 企业。 打造国家级 AI for Science 协同平台，推动国产科学大模型与超算深度适配。 比如让盘古大模型对接大连化物所、高能所的科研需求，形成数据喂模型，模型用算力，算力出成果的正向循环。 让国有存量资源与民营创新活力深度融合。 最后，也是最关键的一点，是坚守自主可控的底线，筑牢底层根基。 从超算架构到 AI 芯片，从并行调度系统到科学大模型，全程立足国产技术，不依赖海外供应链与架构体系，避免走改造即受制的老路，让算力攻坚的每一步都踩在自主可控的土地上，既破解自身短板，又规避卡脖子风险。 当年诸葛亮拆解曹操83万大军，终助孙刘联兵。 赤壁破敌，今日我们拆解美国创世纪计划的虚实，并非为了冷眼旁观，而是为了找准方向，精准发力。 美国的存量超算改造是财力与能源受限下的务实之选，中国的算力攻坚是直面短板、自主创新的必然之路。 无需被美国的宣传声势吓到，亦不能忽视其底层逻辑的可取之处。 战略上保持定力，战术上精准施策，盘活自身存量资源，打通协同壁垒，坚守自主底线，便能让中国的算力优势真正转化为科研优势、产业优势。 在这场全球算力攻坚的棋局中，走出属于我们自己的制胜之路。
英文翻译
The Longzhong Strategy of Computing Power: Unraveling the Reality Behind America’s Genesis Project and Understanding the Core Chess Game of the Sino-US Computing Power Struggle. The world knows that Cao Cao led 830,000 troops south of the Yangtze River, his banners blotting out the sun, striking fear into Sun Quan and Liu Bei. Only Zhuge Liang saw through the reality at a glance—what was called a million-strong army was merely over 100,000 elite troops from Qingzhou, 200,000 recruited naval forces, and the rest were common laborers and conscripts padding the numbers. It seemed mighty but had few true elites. Today, the Genesis Project, grandly launched by the United States, is akin to Cao Cao’s southern campaign back then: 140PB of massive data, collaboration among 24 top-tier enterprises, and a resounding call for AI-empowered frontier scientific research, sparking intense public debate. But if we dissect it thread by thread, like Zhuge Liang, we will see its hype outweighs its substance. The real hard-hitting move lies hidden behind the gimmicks: the transformation of existing supercomputing capacity—exactly like Cao Cao’s elite forces. Strategically, we should despise its pretense; tactically, we must take its substance seriously. We need to identify our own weaknesses and break through precisely, neither deceived by its bluster nor trapped by the gap. The emptiness of the Genesis Project is hidden within its layered promotional narrative, echoing the same logic of Cao Cao’s inflated 830,000 troops. First, the 140PB of data is a bookkeeping bubble—like the common laborers in Cao Cao’s army, possessing bulk but no core value. This accumulated data, spanning decades, consists of redundant records from past experiments in U.S. national laboratories: abandoned data and historical remnants with substandard precision. The vast majority is data waste with no scientific value. The high-value new data and curated old data that could truly support cutting-edge research—such as nuclear fusion or quantum physics—amounts to only a fraction. The so-called full-scale mining is merely a promotional tactic to build hype and secure funding, no different from Cao Cao using laborers to inflate his numbers. Second, the collaboration of 24 enterprises is a lightweight, symbolic maneuver, with no binding commitments or significant capital invested. Companies like OpenAI, DeepMind, IBM, and Oracle have signed only a memorandum of understanding—essentially a policy-oriented PR move to position themselves favorably. They are opening existing models and sharing cloud resources, not injecting tens of billions in dedicated investment. Their core teams remain focused on their own frontier R&D, and mining old data is just an intern-level side task. It is like Cao Cao’s recruited naval forces: many in number, but uneven in combat effectiveness, hardly a core force for breakthroughs. Third, government funding is more lip service than substance, with strained finances unable to sustain the grand plan. The U.S. government’s officially announced initial investment is only $320 million, far from the previously rumored $850 billion total. Compounded by mounting national debt and budget cuts to the Department of Energy, follow-up funding is completely uncertain. This so-called national-level assault is nothing more than political fireworks released before the Christmas holiday—a lot of noise, then nothing but a mess. Peeling away the promotional fog, the substance of the Genesis Project lies in the hard-core layout of transforming existing supercomputers. This is the key variable the U.S. government can still leverage under dual constraints of financial and energy limits, and also the core lever for companies like Nvidia to break out of their predicament. The U.S. currently faces two intractable dilemmas. First, the energy bottleneck of AI computing power is becoming acute. Estimates suggest that by 2026, the power gap for new AI computing centers will reach 30%–40%, with regions like Texas and Arizona already issuing power shortage warnings. GPU clusters consuming over one billion kilowatt-hours per year have become unsustainable energy hogs. Second, companies like Nvidia face stagnant incremental markets. With China’s market largely closed due to export restrictions, the global demand for incremental computing power is unsustainable, leaving only the existing stock market as the sole growth curve. Under these circumstances, retrofitting traditional supercomputers at national laboratories becomes a logical win-win strategy. Supercomputers at top labs like Oak Ridge and Lawrence Livermore are computing hubs built with massive national investment, supported by dedicated nuclear and hydropower, with energy costs far lower than commercial data centers. They already possess solid large-scale parallel computing foundations but previously lacked AI heterogeneous computing capacity. Now, using chips like Nvidia’s GB200 for AI-oriented upgrades essentially installs an intelligent brain onto existing high-quality computing capacity. This avoids building new high-energy-consuming clusters (solving the energy dilemma), quickly fills the AI computing gap in national labs, and allows Nvidia to pivot from selling incremental clusters to transforming old stock, hedging against market decline. This binding of the government revitalizing existing resources and enterprises mining existing stock markets is the true core of the Genesis Project. Seemingly low-key, it precisely targets the U.S.’s current pain points—a pragmatic and efficient way to break through. After seeing through the reality and facade of the U.S. project, we know we must uphold the core principle of despising strategically and respecting tactically: despise the bluff of its propaganda, respect the underlying logic of merging existing supercomputers with AI, and, more importantly, face our own weaknesses and difficulties head-on, forging a self-controlled path for computing power breakthroughs. Admittedly, China also has prominent weaknesses in computing power coordination and supercomputer retrofitting. First, the parallel computing system of traditional supercomputers is cumbersome and outdated—a roadblock for both science and industry. Anyone who has done MPI programming earlier knows that traditional supercomputers are CPU-centric and rely on complex proprietary programming systems for parallel scheduling, making the barrier far beyond the capability of ordinary programmers and researchers. The peak performance of many top-tier supercomputers ends up as numbers on paper, difficult to translate into actual scientific results. Second, supercomputers, AI enterprises, and national laboratories are fragmented, forming information silos. We are not short of top supercomputers—Sunway, Tianhe series; not short of leading AI enterprises—Huawei, ByteDance, etc.; and not short of strong national labs. But these three have long operated independently. National labs have data and supercomputers but lack easy-to-use AI models; AI enterprises have models and AI computing power but lack core scientific data and supercomputer support. Computing power, data, and models cannot form a closed loop, preventing the synergistic effect of 1+1>2. Third, blindly imitating the U.S. path could easily lead to a bottleneck trap. Nvidia’s chip monopoly and the technical barriers of foreign architectures are Damocles’ swords hanging over our heads. Facing the U.S. layout and our own weaknesses, China’s countermeasures must be precisely targeted while rooted in self-reliance—no one-size-fits-all approach, no drifting with the tide. We must focus on core pain points and advance steadily. The first priority is to tackle the parallelization transformation of traditional supercomputers, solving the core problem of being cumbersome and difficult to use. The core of the transformation is not simply adding AI chips, but reconstructing the underlying architecture and lowering the threshold. Use domestic AI chips—Huawei Ascend, Hygon Deep Computing, etc.—as the core to build a heterogeneous parallel architecture of CPU plus domestic AI chips, restructure the underlying scheduling system to achieve efficient coordination between high-precision scientific computing and high-throughput AI computing. Simultaneously build a low-code parallel programming platform, allowing ordinary researchers to easily access supercomputing and AI computing power without mastering complex parallel algorithms, making top-tier computing power truly usable, easy to use, and fruitful. Second, open up the coordination channel linking supercomputers, national labs, and AI enterprises, breaking information silos. With government coordination, establish a hierarchical data sharing mechanism, opening high-value non-classified scientific data from national labs to domestic AI enterprises in compliance. Build a national AI for Science collaboration platform, promoting deep integration of domestic scientific foundation models with supercomputers. For example, let the Pangu model connect with the research needs of the Dalian Institute of Chemical Physics and the Institute of High Energy Physics, forming a virtuous cycle: data feeds the model, the model uses computing power, and computing power yields results. Deeply integrate state-owned stock resources with private innovation vitality. Finally, and most crucially, adhere to the bottom line of self-control, solidifying the foundational base. From supercomputer architecture to AI chips, from parallel scheduling systems to scientific foundation models, the entire process must rely on domestic technology, not dependent on foreign supply chains or architectures, avoiding the old path where transformation equals being constrained. Let every step of the computing power struggle tread on the solid ground of self-control, solving our own weaknesses while avoiding risks of being blocked. Back then, Zhuge Liang dissected Cao Cao’s 830,000 troops and helped Sun and Liu form an alliance to defeat Cao at Red Cliffs. Today, we dissect the reality and facade of America’s Genesis Project—not to stand by indifferently, but to find our direction and exert precise force. The U.S.’s retrofitting of existing supercomputers is a pragmatic choice under financial and energy constraints. China’s computing power struggle is an inevitable path of facing weaknesses and independent innovation. Don’t be intimidated by the U.S. propaganda, nor ignore the merits of its underlying logic. Maintain strategic composure, implement precise tactics tactically, revitalize our own stock resources, break through coordination barriers, and hold the bottom line of self-reliance. Only then can China’s computing power advantage truly transform into a scientific research and industrial advantage, forging our own path to victory in the global chess game of computing power.
back to top