UK’s Cambridge AI supercomputer Dawn shuts down for a week amid 37.7°C heatwave

The recent news about Europeans frantically buying Chinese air conditioners has been very popular, but is there a possibility that:

What needs air conditioning the most? AI supercomputers. (doge)

Far away in the UK, something like this happened these days:

One of the UK's most powerful AI supercomputers, Dawn, shut down completely for an entire week in temperatures exceeding 30 degrees Celsius.

This supercomputer, located at the University of Cambridge, carries an impressive set of labels:

As a core component of the UK government’s £300 million National AI Computing Initiative, 1,024 Intel GPUs and 256 liquid-cooled servers have supported over 350 research projects.

In January this year, it received a £36 million upgrade to scale its capacity, with performance expected to increase sixfold.

Then, at the end of June, a heatwave hit, and it shut down.

More astonishingly, the research running on this supercomputer includes climate change simulations.

What? A machine used to predict global warming has been defeated by global warming.

Here's what happened.

In June this year, the UK experienced its hottest June heatwave on record.

On June 26, the town of Linwood in the Norfolk region reached 37.7°C, breaking the previous June record of 35.6°C set in 1957 and 1976.

The UK Met Office has issued a rare three-day red alert for extreme heat.

Over a thousand schools suspended classes, railway signals failed due to high temperatures, and road surfaces began to melt.

Then, on June 27, as that day’s heatwave peaked, the cooling system at the Cambridge West data center, where the Dawn supercomputer is located, failed to cope.

(P.S. Lincoln and Cambridge are both in eastern England, approximately 103 kilometers apart.)

Dawn, stop here.

△

After the incident, a spokesperson from the University of Cambridge stated to the public:

Dawn experienced technical issues during the heatwave; cooling capacity has been fully restored, and access is expected to resume on July 6.

No specific reason was given, but the situation is:

From June 27 to July 6, Dawn was completely "cooled down" for over a week.

For a supercomputer that burns money every hour and advances scientific research every second, this week’s shutdown was truly terrifying.

Here it is—the one hurt the most has already been revealed.

Professor Vendruscolo’s team at the University of Cambridge is using Dawn to screen for new drug molecules for Parkinson’s disease.

Dawn's machine learning capabilities can screen billions of molecules within days to find compounds that bind to protein aggregates associated with Parkinson's.

If using traditional methods? It would take at least six months, cost millions of pounds, and still only cover a small fraction of what Dawn can scan in a few hours.

A one-week shutdown means this lifeline production line has come to a complete halt.

Lennard Lee from the University of Oxford, lead of the UK’s Cancer Vaccine AI and Supercomputing Project, has been allocated 10,000 GPU hours on Dawn to accelerate AI-driven discovery of targets for personalized cancer vaccines.

Lee previously said:

Discoveries that once took years to achieve now take only a few weeks.

Although Lee later stated that no data was lost and no work needed to be redone, his tone of relief itself underscores the seriousness of the issue.

Additionally, the British Antarctic Survey paused its IceNet sea ice prediction model trained on Dawn, and Cambridge PhD student Bill McGough halted his AI project for kidney cancer screening trained on Dawn... Nearly all of the more than 350 projects running on Dawn were shut down.

Yet all of this was caused by nothing more than 37.7°C.

Alright, the "culprit" has been caught—so who should ultimately be held responsible?

After going around in circles, it seems no one wants to take the blame.

Dawn's cooling system is supplied by USystems, a subsidiary of the French Legrand Group. Following the incident, USystems issued a statement:

Our equipment operated normally and exactly as specified throughout the event.

The cooling system failed, but it's not our fault—our equipment wasn't designed to operate at this temperature.

Is it that the design standards are too conservative, or is climate change happening too quickly?

The answer could be: both.

The historical extreme temperature in the UK in June was only 35.6°C, so Dawn’s cooling system was most likely designed for a similar magnitude.

37.7°C, it's above the limit.

Moreover, this "exceedance" came without warning, as the last time this record was reached was nearly 50 years ago.

Moreover, Dawn is not the only victim.

In the same week, Queen Alexandra Hospital in Portsmouth, UK, declared an emergency due to a cooling unit failure.

The operating room is closed, the catheterization lab is closed, and the imaging department is closed. The hospital informed patients:

Please bring plenty of drinking water, as the hospital is very hot.

Norfolk and Norwich University Hospital (NNUH) fared even worse:

The cooling systems of all MRI scanners failed due to high temperature and humidity, resulting in at least 254 outpatient appointments being canceled.

So, in a sense:

It's not that the supercomputer is fragile—it's that the UK's entire climate control infrastructure wasn't prepared for this kind of weather.

The fact that Dawn was overwhelmed is not surprising at all when viewed over a longer time horizon.

In July 2022, the UK experienced its hottest day on record (40.3°C).

The cooling system at Google's London data center experienced a simultaneous failure of multiple redundant systems, forcing a shutdown to protect the hardware. Google Cloud services in the London region were interrupted for over 18 hours before full restoration.

△

Oracle's London South data center went offline on the same day, and Oracle's statement used an interesting term: "unseasonable heat."

From 2022 to 2026, four years have passed, and a similar event has occurred again.

I just want to ask—how is this issue so difficult that it can’t be prevented in advance?

In reality, causing a supercomputer to crash at temperatures above 30 degrees does have some merit—the most difficult bottleneck to overcome is heat dissipation.

In particular, for the European region, equipment commonly uses natural cooling, which is inherently limited by outdoor ambient temperatures.

How to understand it?

All cooling systems, no matter how advanced, ultimately transfer heat to outdoor air, which is the ultimate bottleneck for the entire chain.

This link expands as follows:

The chip transfers heat to the heatsink, the heatsink transfers it to the coolant or air, the coolant transfers it to the cooling tower, and the cooling tower releases it into the atmosphere.

The market is the last one to hold the bag.

So when the atmosphere itself reaches 37°C, it starts to lose its ability to cope.

Specifically, when outdoor temperatures rise from 20°C to 37°C, the cooling efficiency of cooling towers and dry coolers may drop by 40% to 50%.

You asked why the air conditioner isn't on? Because the compressor's efficiency decreases and current increases at high temperatures, making it prone to overheating and tripping.

The 2022 Oracle incident report stated verbatim: "Two cooling units failed when required to operate beyond their design limits."

Dawn's situation this time is also reasonably assumed to be similar.

It uses Dell PowerEdge XE9640 servers equipped with a direct liquid cooling system, a far more advanced cooling solution than traditional air cooling.

Coolant flowing directly over the chip surface removes heat far more efficiently than air cooling.

But as always, liquid cooling addresses efficiency within the rack. After heat is carried away by the coolant, it must still pass through a cooling distribution unit, facility chilled water loop, and cooling tower before being released into the outdoor atmosphere. The final step remains constrained by outdoor temperatures.

Once the cooling system stops, it will trigger a series of subsequent chain reactions.

Research data shows that once the cooling system shuts down, the server inlet temperature can rise from 22°C to over 35°C within five minutes.

In this situation, the chip will activate its self-protection mechanism:

First, thermal throttling reduces clock speed to lower heat generation, causing a significant drop in performance; if the temperature continues to rise and exceeds the safety threshold, the system will shut down forcibly.

At this point, the operator has only two options:

Allowing the device to power off automatically may corrupt data;

Perform an orderly shutdown to protect hardware, but business operations will be suspended.

Google, Oracle, and Cambridge Dawn all chose the latter.

There is something even more concerning.

As AI data centers continue to expand, the impact of temperature on AI may become increasingly significant.

A few days ago, I watched Xiao Lin's on-site visit to Huawei's data center on Bilibili, and one comparison left a strong impression:

Traditional data center racks typically have a power density of about 5 to 10 kilowatts, but AI training racks have reached 30 to 50 kilowatts; Nvidia’s latest GB200 NVL72 rack has surged to 120 to 132 kilowatts (with the next-generation Rubin potentially reaching 600 kilowatts).

What does that mean? An AI rack generating 100 kW of heat is equivalent to running 50 space heaters simultaneously in a space the size of a phone booth.

Imagine all the small heaters used in winter packed into a single rack—that’s the cooling challenge faced by today’s AI computing infrastructure.

△

Worse still, GPUs themselves are becoming increasingly "hot."

The Nvidia V100 in 2017 consumed about 300 watts; the H100 in 2023 increased to 700 watts; the B200 in 2024 reached 1,000 watts; and the B300 and AMD MI355X in 2025–2026 are set to jump directly to 1,400 watts.

Over seven years, the heat generation of a single chip has increased three to five times.

So, whether in terms of quantity or individual chips, as AI becomes more powerful, it generates more heat and requires better cooling.

At this point, we can see two curves colliding:

The chip is heating up exponentially, and the Earth is also warming faster.

Things are starting to get more complicated.

Google built a data center in Finland as early as 2011, and Meta went to northern Sweden—to use the cold climate for natural cooling.

Elon Musk even thought about building an AI data center in space.

But in January this year, the UK government just granted Dawn £36 million to expand its capacity, and is also planning a new national supercomputer in Edinburgh.

It’s unclear whether the cooling design for these facilities was based on last era’s British summers or the new normal that’s just beginning.

But one thing is certain:

The supercomputer used to predict climate change was shut down by heat from climate change.

This is no longer a joke; it's a real challenge facing the infrastructure of the AI era.

Reference link:

[1]https://www.thetimes.com/uk/science/article/cambridge-ai-supercomputer-heatwave-shutdown-ns7rcmkgs

[2] https://www.datacenterdynamics.com/en/news/data-center-housing-uks-dawn-supercomputer-suffers-heatwave-related-outage-report/

[3]https://x.com/cashandcarrots/status/2074016783812505762

This article is from the WeChat official account "Quantum Bit," authored by Yi Shui.