Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time β€” we’ll notify you if affected. OEM costs and supply tightening β€” contact us to secure stock and pricing. For projects or bulk orders, choose β€œRequest for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Servers & Power

Fanless Supercomputers: What 100% Direct Liquid Cooling Looks Like, and What You Give Up

Walk up to a typical liquid-cooled AI rack and you will still hear it. Cold plates take the heat off the GPUs, but drives, memory, power electronics and network cards are cooled the old way, by fans.

A fanless supercomputer is different. There is nothing spinning. Every heat source, down to the power rectifiers and network switches, is cooled by liquid. HPE calls this 100% fanless direct liquid cooling and uses it in its Cray Supercomputing systems.

It is an elegant piece of engineering. It also means giving up something data centres have always had: air as a fallback.

Hybrid liquid cooling vs fanless

What πŸ’§πŸ’¨ Hybrid (e.g. NVL72 racks) πŸ’§ 100% fanless (e.g. Cray EX)
Heat removed by liquid About 90%
All of it
Heat left for the room's air About 10%
Essentially none
Fans inside βœ… Yes, fewer ❌ None
What is liquid-cooled CPUs, GPUs, some switches CPUs, GPUs, memory, rectifiers, switches
πŸ”Š Noise 🟑 Reduced 🟒 Very low
Room air conditioning still needed βœ… Yes 🟒 Minimal
If the liquid loop stops πŸ”΄ System stops πŸ”΄ System stops

Sources: HPE GB300 NVL72 for the 90/10 split, Data Center Dynamics on HPE's fanless architecture

What HPE claims for it

Claim Figure Compared with
⚑ Cooling power per server blade 37% lower Hybrid direct liquid cooling
⚑ Cooling power consumption Potential for 90% lower Traditional air-cooled systems
πŸ“ Floor space About half Less dense designs
πŸ”₯ Density Up to 400 kW per rack without fans β€”

Sources: HPE press release, October 2024, HPE direct liquid cooling. These are HPE's own figures.

HPE describes the architecture as the industry's first of its kind, and points to its track record: in the November 2024 TOP500 list, seven of the ten fastest supercomputers ran on HPE Cray Supercomputing EX systems using it.

Why removing the last 10% matters

The leftover heat is not small

Ten percent sounds trivial. On a 140 kW rack it is about 14 kW, which is more than many entire enterprise racks produce. Multiply by a row of racks and the hall still needs serious air conditioning for equipment that is supposedly liquid-cooled. Going fanless removes that second cooling system from the bill and from the building.

Fans use power too

Server fans are not free. At high density they consume a meaningful share of a server's power and create most of its noise. That is where HPE's 37% figure comes from: the energy that hybrid systems still spend moving air.

Density stops being limited by airflow

With no air path to preserve, components can be packed far more tightly. HPE quotes up to 400 kW per rack without fans, well beyond what any design that needs air channels could manage.

What you give up

There is no plan B

In an air-cooled hall, a cooling fault usually means things get warm and throttle. There is time. In any liquid-cooled system, hybrid or fanless, loss of the coolant loop stops the machine. Fanless designs take that dependence to its conclusion: nothing in the cabinet has an alternative. The facility's liquid loop, pumps and heat rejection have to be engineered with the redundancy that implies.

Everything has to be built for it

You cannot make a server fanless by adding a kit. Memory, power supplies and switches all need cold plates and plumbing designed in. That is why fanless cooling appears in purpose-built supercomputer cabinets and not in general-purpose servers. It also ties you more closely to one vendor's cabinet, blades and service organisation.

Servicing is different

Every component swap involves coolant connections. The designs use dripless couplings, and the people doing the work need training. It is routine in supercomputing centres and unfamiliar almost everywhere else.

The headline numbers are the vendor's

The 37% and 90% figures are HPE's, measured against comparisons HPE chose. They are plausible and consistent with the physics. They are not independent benchmarks, and the saving in your facility depends on what you are comparing with.

Is this coming to ordinary AI racks?

The direction is clear. NVIDIA has described its next-generation Rubin servers as having sealed front panels and no internal fan noise, a redesign that extends liquid cooling to components earlier systems left to air. What was a supercomputing speciality is becoming the mainstream design for flagship AI hardware. We look at that shift in The Last Air-Cooled Flagship?

Source: DataCentreNews on Rubin servers

Who should care today

Your situation What fanless means for you
πŸ›οΈ Building a supercomputing facility βœ… The reference design. Plan the building around it.
πŸ—„οΈ Deploying NVL72-class racks 🟑 You are on hybrid. Budget for the air-cooled remainder.
πŸ—οΈ Designing a new data hall for the next decade 🟑 Assume future flagships need little or no air. Size the liquid plant accordingly.
πŸ’¬ Running inference servers in a normal hall 🟒 Not your concern yet.

A closing thought. Seymour Cray liquid-cooled a supercomputer in 1985. The idea is forty years old. What is new is that the rest of the industry has run out of alternatives.

Related on SourceIT

Share This Article
Facebook

More from our blog


← Older Post Newer Post β†’