NVIDIA GTC Berlin Finale: Hermes Agent Bridges the Gap to Robot Deployment

In a supercomputer facility where tens of thousands of GPUs are linked into one giant cluster, what drives operations engineers to despair has never been chips that compute too slowly, but rather the spiderweb of cables with poor contacts and the dead zones in cooling. The Vera Rubin node that Jensen just held up on stage tackles both of these pain points at the root with the most brutal industrial design: zero cables and zero wires inside, 100% full liquid cooling, and a node that used to take 2 hours to screw together can now be installed in under 5 minutes. On stage, he turned this new node, fitted with two GPUs, around to show it off—inside there was not a single cable or wire—and it went from the previous generation's 80% liquid cooling all the way to complete 100% full liquid cooling. Outsiders watching NVIDIA's launch events usually fixate only on how many times this year's per-card compute has grown and how many more GB of memory there is, always assuming that the compute arms race ultimately comes down to a pure TSMC wafer-process game. But Jensen's move to cut out all cables directly exposes the real bottleneck in the era of ten-thousand-GPU clusters. When big companies start cramming tens of thousands of top-tier GPUs into the same data center, the most lethal source of failure is no longer the chip itself, but the copper cables packed like a spiderweb in the previous generation of nodes and that 20% of cooling blind spots that still relied on fans. Workers had to screw together wiring harnesses for two hours per unit, and once the data center was running, a single loose cable or a bit of accumulated dust could instantly interrupt and restart a training job worth tens of millions of dollars. It is like taking the old workshop-style assembly line—with wires flying everywhere and electric fans blowing hot air—and encapsulating it directly into a perfectly fitted Lego brick. Pulling out all the cables means completely eliminating the physical failure points of manual miswiring and poor contacts, while 100% full liquid cooling wipes out the last remaining air-cooling dead zones, compressing two hours of hard labor into five minutes of standardized blind-mate installation. Put plainly, this reveals an extremely harsh bottom line to everyone watching compute infrastructure, doing hardware engineering, and tracking tech cycles: the moat of top tech giants has long since shifted from single-point chip design to factory manufacturing cadence and data center operations efficiency. Cutting assembly time from 120 minutes to under 5 minutes means the shipment turnover speed of contract manufacturers has more than twentyfolded, and the downtime maintenance costs for cloud providers deploying new data centers have been slashed to the bone. This means third-rate hardware upgrades are about desperately stuffing wires into an old chassis, while a first-rate industrial revolution is about compressing two hours of threading needles into five minutes of Lego-style blind-mate installation. https://t.co/RQc0YOulNd

View original

在几万张显卡连成片的超级机房里，最让运维工程师崩溃的从来不是芯片算得不够快，而是蜘蛛网一样的线缆接触不良和散热死角， 老黄刚在台上举起的这台 Vera Rubin 节点，直接用最暴力的工业设计把这两个痛点连根拔起：内部零电缆、零电线、100% 全液冷，原本要拧 2 小时的节点现在不到 5 分钟直接装完。 他在现场把这台装了两块 GPU 的新节点翻过来展示，里面竟然连一根电缆、一根电线都没有，而且从上一代的 80% 液冷直接做到了彻底的 100% 全液冷。 平时外行看英伟达的发布会，默认只盯着今年单卡算力又涨了几倍、显存又多了多少 G，总觉得算力军备竞赛拼到最后纯粹是台积电的晶圆制程游戏。 但老黄这把砍掉所有线缆的操作，直接点破了万卡集群时代真正卡脖子的死穴，当大厂开始把几万张顶级显卡塞进同一个机房时， 最致命的故障源早就不是芯片本身了，而是上一代节点里蜘蛛网一样密布的铜缆和那 20% 还要靠风扇吹的散热盲区，工人装一台要拧两小时线束，机房跑起来只要松一根线或者积一点灰，几千万美金的训练任务就得当场中断重来。 这就像把以前满地飞线、还要靠电风扇吹热气的作坊式流水线，直接封装成了一块严丝合缝的乐高积木，拔掉所有线缆意味着彻底消灭了人工接错线和接触不良的物理故障点，100% 全液冷则把最后一点风冷死角彻底清零，把两小时的苦力活压缩成五分钟的标准化盲插。 说白了，这给所有在盯算力基建、做硬件工程和看科技周期的兄弟透了一个极狠的底牌，顶级科技巨头的护城河早就从单点芯片设计，彻底卷到了工厂制造节拍和机房运维效率上，组装时间从 120 分钟砸到 5 分钟以内，意味着代工厂的出货周转速度直接翻了二十多倍，云厂商部署新机房的停机维护成本更是被砍到了脚脖子。 这意味着三流的硬件升级是在旧机箱里拼命塞满电线，一流的工业革命是把两小时的穿针引线直接压成五分钟的乐高盲插。 https://t.co/RQc0YOulNd