惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

D
Docker
阮一峰的网络日志
阮一峰的网络日志
T
Tailwind CSS Blog
博客园 - 【当耐特】
量子位
博客园 - 叶小钗
有赞技术团队
有赞技术团队
Jina AI
Jina AI
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
博客园 - Franky
博客园 - 司徒正美
爱范儿
爱范儿
美团技术团队
小众软件
小众软件
酷 壳 – CoolShell
酷 壳 – CoolShell
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
V
V2EX
罗磊的独立博客
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
Last Week in AI
Last Week in AI
Hugging Face - Blog
Hugging Face - Blog
I
InfoQ
D
DataBreaches.Net
宝玉的分享
宝玉的分享

Blog – Hackaday

Why Not Yserver? It’s Xserver, But Rust-y. OpenCAL: Computed Axial Lithographic 3D Printing For Everyone Is A CS Degree DOA Thanks To LLMs? IEEE Says TBD. Double The VRAM Of An RTX 3070 The Pacemaker Patch Robot Chess But Each Piece Is A Small Robot Bambuddy Says Bye To Bambu Lab Cloud Services Converting A Scanning Electron Microscope Into A TEM Is Surprisingly Easy Custom Watch Is On The Case Patterns Everywhere Behold A 60 Hz Refresh Rate E-ink Monitor GentleOS, A Simple OS For Your Old PC Deeply Optimized MSX Emulation On ESP32-S3 With VGA Output Homebrew Macropad Looks Good The Air Position Indicator For The B-29 Building A 1:150 Scale Toyota ProBox Micro Remote Control Car Adding Weight To A 3D Print With Plaster Of Paris, Cleanly Hackaday Podcast Ep 373: GPS, Danger In Space, And Robby The Robot A Peek Inside The Secret Lagercrantz Suitcase Radio This Week In Security: Microsoft On Microsoft, Register Your Domains, Linux On ARM, And FreeBSD Joins The File Cache Club Glue-in Hinge Design Tries Something Different The Hackaday Communicator Badge, Re-Imagined With New Firmware Amiga 1232 Storm CD Packs Every Upgrade Into One Wedge So Many Analog To Digital Converters Repairing A Pair Of Voodoo 2 GPUs For Some SLI Action AI The Truly Environmentally Friendly Way Evidence For Water Vapor Plumes On Europa Vanishes In Re-Analysis Mechanical Stability For Your Coils 3D Printed Hose Sprayer Sets Phasers To Suds The Merits Of Comment-Driven Development As Counterweight To TDD
NVIDIA’s New AI Servers Run On Hotub Coolant And Don’t Ne...
CncPartner · 2026-06-27 · via Blog – Hackaday

When people start ranting about AI, you can be sure a few things are going to come up during the two-minutes hate: job loss, higher power bills, the neverending tide of low-effort slop, and wasting precious freshwater. Well, NVIDIA wants to take away that last one, beacause the all-water cooled Ruben architecture won’t need any evaporative cooling— coolant can stay in a closed loop, and never needs to be cooled below 45 C, or 113 F.

This sort of coolant loop should be familiar to anyone who has ever built a water-cooled PC or PlayStation: there’s a glycol-water mix, water blocks, and a radiator to reject heat to the environment. NVIDIA doesn’t mention if their new servers come with RGB lighting, but we’d like to imagine it’s an option. The big difference — aside from the rainbow LEDs– between a Ruben server and your old gaming rig is that in these racks, everything is on a waterblock. If there’s a chip on the motherboard generating heat, it’s getting rid of it into the same cooling water. Cooling water, that we have to emphasize, needs only be cooler than the chips themselves: in this case, they’re talking 45 C on the cold side, and 55 C headed out of the racks. (That’s 113 F to 131 F for all the bald eagles reading this.)

Given the required temperature drop is so modest, there’s no need for the evaporative chillers that have given AI data centers such a bad name in water conservation circles. Just like in a water-cooled PC, ambient-temperature air running over dry heat exchangers– also known as big honkin’ radiators–is able to handle the cooling, so no water is lost. Since everything is on waterblocks, there’s no need for cooling air, either, and the server farms need only be air conditioned to the degree required to make them comfortable to work in.

If you think NVIDIA is making this change because they suddenly care about water conservation, think again. The press release makes their motivations very clear: cooling costs money, and running this hot saves a lot of it. We’re talking four mil US a year for a 50 MW hyperscaler. One might suspect that this sort of thermal regime could limit the lifetime of the hard-working NPUs, but since they’ll be obsolete in a few years anyway, that’s not likely a big concern, especially not for NVIDIA.

We’ve actually seen hotter fluids used to cool computers before– coffee, for one. Water cooling also isn’t new in the data center world; we took a look at it a few years back. Things are clearly heating up now, though.