
最新一期的超级计算机Top500榜单出炉,但变化不大。中国在半年度排名中拥有最强大的超级计算机,地位无可撼动,美国远远落后。中国的“LineShine”超级计算机位居榜首,根据高性能Linpack(HPL)基准测试,其性能达到2.2艾浮点运算每秒。这大约比美国最大的“铁器”——劳伦斯利弗莫尔国家实验室的“El Capitan”超级计算机——高出22%。
A new Top500 list of supercomputers has dropped, but little has changed. China remains uncontested with the most powerful supercomputer on the biannual ranking, and the US is not even close behind. The Middle Kingdom's LineShine super tops the list, boasting 2.2 exaFLOPS as measured by the High Performance Linpack (HPL) benchmark. That's roughly 22 percent ahead of Uncle Sam’s biggest iron, Laurence Livermore National Laboratory’s El Capitan.
但LineShine未必是最节能的。LineShine使用42.2兆瓦的电力,其领先地位是以明显更高的电费为代价的;不过这似乎并不重要。排名第一就是第一,更让人雪上加霜的是,深圳国家超级计算中心(NSCS)在没有美国GPU可用的情况下做到了这一点,因为美国出口管制切断了它们获取美国GPU的途径。整台超级计算机由45,360颗“LingKun LX2”中央处理器驱动,每颗拥有304个Arm v9.2核心,主频为1.5吉赫兹。
But LineShine may not be the most efficient. Using 42.2 megawatts, LineShine’s lead comes at the expense of a noticeably higher power bill; not that that makes much of a difference. Number one is number one, and to add insult to injury, the National Supercomputing Centre in Shenzhen (NSCS) did it in spite of US export controls denying them access to American GPUs. The entire super is powered by 45,360 LingKun LX2 CPUs, each with 304 Arm v9.2 cores clocking to 1.5 GHz.
美国短期内不太可能重夺榜首,但也并非没有尝试。根据去年晚些时候宣布的“Genesis Mission”计划,美国能源部正按计划部署至少九台新超级计算机。只是这些机器大多是较小的或专注于人工智能的系统,真正的大型系统寥寥无几。橡树岭国家实验室(ORNL)与HPE和AMD合作建造的“Lux”超级计算机就是其中一例。
The US is unlikely to reclaim the number one spot any time soon, though not for lack of trying. Under the Genesis Mission announced late last year, the Department of Energy is on track to deploy at least nine new supers. It just so happens that most of these will be smaller or AI-focused machines, with few truly large systems. Oak Ridge National Laboratory’s (ORNL) Lux supercomputer, built in collaboration with HPE and AMD, is one such example.
在HPL测试中,它提供220拾浮点运算每秒的双精度算力,无疑是一台强大的系统。两年前它还能进入前十,但这次只能在秋季的Top500排名中排在第14位。现在,如果你愿意把精度滑块调低到FP4,峰值性能将跃升至20艾浮点运算每秒以上。这对于人工智能推理甚至部分训练来说足够了,但对于运行美国核威慑的多物理模拟,能源部传统上会选择更强大的系统。ORNL即将推出的“Discovery”系统,肯定有足够的算力将LineShine赶下宝座。
Delivering 220 petaFLOPS of double-precision grunt in HPL, it’s certainly a potent system. Just two years ago it would have been a top 10 supercomputer, but it''l have to settle for number 14 on this fall’s Top500 ranking. Now, if you’re willing to turn the resolution slider down to FP4, peak performance jumps to somewhere north of 20 exaFLOPS. That’s fine for AI inference and even some training, but for running multi-physics sims on America's nuclear deterrent, the DoE has traditionally gone for something a bit bigger. ORNL’s upcoming Discovery system will certainly have enough grunt to give LineShine the boot.
关于基于HPE Cray的系统细节仍然不多,但我们知道,它将由AMD Venice CPU和Instinct MI430X GPU组合驱动。与当今销售的大多数GPU不同,MI430X没有为了低精度AI吞吐量而放弃双精度算力。据估计,这些多芯片组件中的每一款都能提供超过200万亿次FP64矩阵运算。根据能源部的说法,Discovery的性能将是Frontier的3倍到5倍,这将使其FP64峰值性能达到5.1到8.5艾浮点数之间。
Details on the HPE Cray based system remain thin, but we do know it’ll be powered by a combination of AMD Venice CPUs and Instinct MI430X GPUs. Unlike most of the GPUs sold today, the MI430X hasn’t forsaken double-precision grunt for low-precision AI throughput. Each of these multi-die chips is estimated to deliver north of 200 teraFLOPS of FP64 matrix math. According to the DoE, Discovery will deliver somewhere between 3x to 5x the performance of Frontier, which would put its Rpeak somewhere between 5.1 and 8.5 exaFLOPS at FP64.
尽管这听起来令人印象深刻,但中国已经证明,一旦排除电力因素,要夺得第一名只需要规模。而LineShine擅长的正好就是规模。在像El Cap这样的GPU加速系统中,你很难让HPL达到系统理论峰值的70%。但LineShine不是GPU加速系统。当然,拥有16个HBM堆栈的LX2看起来很像GPU,但它实际上只是一台拥有足够向量和矩阵单元的巨型CPU,足以让不少现代GPU相形见绌。
As impressive as that might sound, China has proven that once you take power out of the equation, all it takes to claim the number one spot is scale. And if there is one thing LineShine does well, it’s scale. In GPU-accelerated systems like El Cap, you’re lucky to push HPL to within 70 percent of the system’s theoretical roofline. But LineShine isn’t GPU-accelerated. Sure, with 16 HBM stacks, the LX2 looks a lot like a GPU, but it's really just an enormous CPU with enough vector and matrix units to put more than a few modern GPUs to shame.
只要数量足够——我们指的是22,680个双插槽节点——国家超级计算中心就首次在近十年里将中国推上了Top500榜首。不再是过去的基准测试 美国并不缺乏能够超越El Capitan甚至LineShine性能的系统,只要有人有足够动力在它们上运行HPL。SpaceXAI在田纳西州孟菲斯的300兆瓦Colossus超级计算机拥有20万个Hopper GPU,每个可提供34到67万亿次(向量/矩阵)FP64算力。
With enough of them – by which we mean 22,680 dual-socket nodes – NSCS catapulted China to the top of the Top500 for the first time in nearly a decade. Not the benchmark it used to be The US has no shortage of systems capable of eclipsing El Capitan or even LineShine’s performance if someone were motivated enough to get HPL or running on them. SpaceXAI’s 300 megawatt Colossus supercomputer in Memphis, Tennessee, boasts 200,000 Hopper GPUs each good for 34-67 (vector/matrix) teraFLOPS of FP64 grunt.
即使假设效率低劣的25%,该集群也足以在HPL中突破三艾浮点数。它可能需要El Capitan 10倍的电力,但问题是,除了赢得白宫的青睐——在DOGE事件之后,这艘船对埃隆·马斯克来说几乎肯定已经开走了——将集群租给Anthropic来训练和运行Claude,是一项利润丰厚得多的事业。
Even assuming an abysmal 25% efficiency, the cluster should be more than capable of breaking three exaFLOPs in HPL. It might need 10x the power of El Capitan, but the thing is, beyond winning points with the White House — a ship that has almost certainly sailed for Elon Musk after the DOGE fiasco — renting out the cluster to Anthropic to train and run Claude is a far more lucrative endeavor.
同样的情况也适用于大量人工智能集群,从 Meta 的 Prometheus 到 OpenAI 的 Starship。过去,登顶 Top500 的系统曾是人类建造的最大型计算机之一。但这已不是什么新鲜事。美国急需算力。HPL 甚至其姊妹版高性能梯度共轭基准测试,已不再是需要击败的基准。新的战场不再是核扩散,而是人工智能模型、智能体以及驾驭它们的手段。
The same could be said of any number of AI clusters, from Meta’s Prometheus to OpenAI’s Starship. It used to be that the systems to summit the Top500 were among the largest computers humanity has ever built. But this hasn’t been true in some time. The US is starving for compute. HPL or even its sibling High-Performance Gradient Conjugate just aren’t the benchmarks to beat any longer. The new battleground isn’t nuclear proliferation; it’s one of AI models, agents, and the means to harness them.
我们谈论的不是代币、大语言模型或聊天机器人,而是利用这些模型自动化实验室,并在服务的科学家下班回家吃饭后很久,仍能制定并检验假设的能力。这正是美国“创世”任务如此强调人工智能的原因。人工智能驱动的实验室为研究人员提供了更快迭代的机会。与其运行成百上千次模拟来大海捞针——这类任务确实需要更强大的算力——人工智能分析和自动化可以聚焦这项搜索。
We’re not talking about tokens, LLMs, or chatbots, but the ability to harness these models to automate laboratories and formulate and test hypotheses long after the scientists they serve have gone home for dinner. It’s why the US’s Genesis mission puts such a strong emphasis on AI. AI-driven labs present an opportunity for researchers to iterate faster. Rather than running hundreds or thousands or simulations trying to find a needle in a haystack, a task for which a bigger iron is warranted, AI analysis and automation could focus that search.
只要更高的浮点运算速度意味着单位时间内更多的模拟,高性能计算(HPC)以及由此延伸的 Top500 将保持相关性。但这些系统的运行速度,只能快到操作它们的男女科学家所能达到的程度。®
High-performance computing (HPC), and by extension the Top500, will remain relevant as long as more FLOPS means more simulations per unit of time. But these systems can only move as fast as the men and women who operate them. ®