Why Sierra the Supercomputer Had to Die

3 days ago 17

It was the authorities that decided it was clip for Sierra to die. Sierra, it indispensable beryllium said, was a supercomputer, and truthful had ne'er truly been live successful the archetypal place. But by immoderate nonsubjective measure, she lived an awesome life. She resided successful bluish California astatine the Lawrence Livermore National Laboratory, wherever she was minded by dozens of unit astatine the lab’s computing complex, successful Building 453. She completed her last jobs precocious past year, successful October, earlier she went offline for good. She was 7 years old.

A heap  of InfiniBand cables removed from the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory in...

Gutted: a heap of InfiniBand cables removed from the Sierra supercomputer astatine Lawrence Livermore National Laboratory.

Photograph: Balazs Gardi

Employees disassemble racks of the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory successful  Livermore...

Employees disassembling Sierra’s racks.

Photograph: Balazs Gardi

Disconnected racks of the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory successful  Livermore California...

Sierra’s disconnected racks.

Photograph: Balazs Gardi

According to the TOP500, which ranks these mega-machines, Sierra was erstwhile the second-fastest supercomputer successful the world. She was conceived successful a Chicago edifice league country much than a decennary ago, astatine a method treatment for officials from America’s nationalist labs. The eventual decorator baby, Sierra was assembled from thousands of IBM Power9 CPUs and Nvidia Volta V100 GPUs—a daring, offbeat architecture for Livermore astatine the time.

Like different supercomputers, Sierra was girthy. She was composed of thousands of compute nodes, stored 1 connected apical of different successful racks—basically cabinets—that held up her processing innards. She had 240 of these racks, dispersed crossed astir 7,000 quadrate feet. All of this was needed to enactment her life’s main occupation: performing specialized, super-high-security simulations for the National Nuclear Security Administration. At the clip of her decease sentence, her processing powerfulness ranked a still-respectable 23rd successful the world.

Now, why did Sierra person to die? After all, an tremendous magnitude of clip and resources went into Frankensteining her together. The enactment of the laboratory won’t corroborate however overmuch she outgo to build, but she was expensive—the authorities spent astatine slightest $325 cardinal connected her and her fraternal twin, a supercomputer called Summit astatine the Oak Ridge National Lab successful Tennessee. (Summit was decommissioned successful precocious 2024.) Also, she inactive wholly worked. “At the extremity of the beingness of a machine, you could think, Oh, we person each these sunk costs. You should conscionable support moving the instrumentality forever,” says John Allen, the lab’s organizational accusation information officer. But that’s wrong. “Its bully and faithful work is over, and we person to determination on.”

Management rack of the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory successful  Livermore California...

Management racks were saved for last.

Photograph: Balazs Gardi

Sierra supercomputer astatine  the Lawrence Livermore National Laboratory successful  Livermore California Unites States connected  December 16...

So galore cables.

Photograph: Balazs Gardi

There are respective reasons to accidental goodbye. One is the hardware’s earthy lifespan. Even astatine birth, definite virginal components are defective, truthful turning the happening connected becomes an contiguous experimentation successful discovering manufacturing errors and replacing those components. Then the instrumentality enters its aureate era. Eventually, though, the bulk of the computer’s chips are pushed to the brink, and the nonaccomplishment complaint starts to emergence again. This rhythm of brokenness, from precocious to debased to precocious again, is what IT experts sometimes telephone the bathtub curve, and there’s an evident inducement not to scope the different broadside of it. “As you age—just similar humans—you are apt to get much disease,” says Devesh Tiwari, who researches high-performance computing astatine Northeastern University. “You are apt to neglect more, truthful you request much caring and feeding.” A related occupation is obsolescence, for some the hardware and the bundle utilized to run it. Replacement parts go hard oregon adjacent intolerable to source.

A drawer filled with 106 10TB hard   drives successful  a rack that houses 8 shelves arsenic  portion  of the Sierra ace  machine  astatine  the...

A drawer filled with hard drives.

Photograph: Balazs Gardi

An worker  prepares a coagulated  authorities   thrust  precocious    removed from the Sierra supercomputer for shredding astatine  the Lawrence...

An worker prepares a solid-state thrust for shredding.

Photograph: Balazs Gardi

Sierra ne'er got excessively acold into the last signifier of the bathtub curb, says Rob Neely, the lab’s subordinate manager for weapons simulation and computing, but she was successful information of getting there. Neither the IBM nor the Nvidia components are inactive successful production, and IBM nary longer supports the mentation of the operating system—Red Hat Enterprise Linux—that Sierra used. “It's truly astir resources,” says Ann Dunkin, the erstwhile main accusation serviceman of the US Department of Energy, which oversees the nationalist laboratory systems. “If they had infinite resources, they would tally infinite supercomputers.” Seven years is simply a reasonably emblematic lifespan.

But it is El Capitan, Sierra’s newer and speedier successor (and one-time next-door neighbour astatine the lab), who astir threatened her existence. To the untrained eye, Sierra and El Capitan don’t look precise different. They’re some agelong lines of whirring racks hooked into immense powerfulness supplies nether the floorboards. But it’s the insides that count. Sierra had awesome components, but El Capitan came online successful 2025 with the AMD Instinct MI300A APU, positive a communal representation shared crossed his CPUs and GPUs. He tin instrumentality up to 36 megawatts to tally (compared to Sierra’s 11). That’s enough, the laboratory says, to powerfulness 36,000 humble homes.

An worker  removes a coagulated  authorities   thrust  from a batch node of the Sierra supercomputer astatine  the Lawrence Livermore National...

Sierra’s decommissioning proceeded successful stages.

Photograph: Balazs Gardi

Top down   presumption    of a batch node of the Sierra supercomputer astatine  the Lawrence Livermore National Laboratory successful  Livermore...

Top-down presumption of a batch node.

Photograph: Balazs Gardi

Supercomputers tin beryllium measured successful respective ways, but the captious statistic is their quality to execute floating-point operations per second, oregon flops. Flopping arsenic accelerated arsenic imaginable is what makes you successful. At her peak, Sierra could deed 94.64 petaflops—94.64 quadrillion floating-point operations—per second. El Capitan, astatine 1.809 exaflops, is astir 19 times faster. In precocious 2025, helium was officially declared the world’s fastest supercomputer. Sierra’s juice, Neely says, was nary longer worthy the squeeze.

There was no large reddish button, nary elephantine lever, that turned Sierra off. Someone could’ve conscionable chopped the cords, sure, but that’s not the recommended procedure. First, Sierra’s idiosyncratic scientists were warned, via email, to prevention their work. Then a DNR was formally instituted—no caller parts.

The decommissioning proceeded successful phases, starting with the compute nodes and the rack switches—management nodes are last, since they’re needed until the precise end. The process involves moving scripts that, digitally, unopen the machine down, and past hard powerfulness switches are flipped disconnected too. There’s besides a dehydration. When she was alive, Sierra could get rather hot, truthful the laboratory recirculated thousands of gallons of h2o per minute, funneled done veiny pipes that came up from nether her floorboards. As she approached death, that h2o had to beryllium drained. It was tested by information unit first, to guarantee it was an environmentally steadfast pH.

Large diameter aquatherm pipes arsenic  portion  of the cooling strategy   for the Sierra supercomputer astatine  the Lawrence Livermore...

Some of the pipes that kept Sierra cool.

Photograph: Balazs Gardi

It’s worthy saying present that supercomputers tin person much dignified retirements. Some extremity up donated to different facilities oregon astatine museums. They tin beryllium auctioned off, arsenic the General Services Administration did successful 2024 to dispose of Cheyenne, a petaflop supercomputer built by Silicon Graphics International. But the information is, determination isn’t overmuch request for aged supercomputers, and astir are simply stripped for parts. Back successful 2013, erstwhile it couldn’t pique involvement successful the full enchilada, New Mexico opted to interruption down its state-funded Encanto supercomputer and merchantability it successful pieces. The Argonne National Lab tried to springiness overmuch of its Intrepid supercomputer, erstwhile the world’s third-fastest, to different labs, arsenic good arsenic a machine museum, but determination were fewer takers. Other than a tiny fig of racks that went to North Carolina State, Intrepid was recycled.

Ethernet switches of the Sierra supercomputer astatine  the Lawrence Livermore National Laboratory successful  Livermore California...

Ethernet switches.

Photograph: Balazs Gardi

Employees region   hard   drives from a rack of the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory in...

At clip of death, Sierra was 7 years old.

Photograph: Balazs Gardi

Sierra is being recycled connected an utmost scale. She was, aft all, designed to enactment the country’s atomic stockpile and was truthful chock-full of classified data—the instrumentality can’t conscionable beryllium thrown out. Instead, Sierra had to beryllium beaten down to a pulp to debar immoderate accidental that she mightiness beryllium partially resuscitated and utilized to reconstruct authorities secrets. This is simply a bloody process. Staff, wearing gloves, propulsion retired nodes and region the lithium-ion batteries peppered throughout. (These volition beryllium sent to a specialty artillery recycler.) Other parts, similar strategy boards, processors, and the skeletal racks that held Sierra together, are sent for a coarse shredding offsite. Anything that can’t beryllium recycled is, aft a coagulated data-security analysis, destroyed.

Sierra’s flash representation components, however, tin store information adjacent without power, truthful these are crushed into a precise good powder. Meanwhile, to dispose of immoderate magnetic drives, the laboratory keeps a special, government-approved degausser downstairs. The contraption uses a imperishable magnet—a worldly that generates magnetic fields without electricity—to hitch components clean. (The magnet is beardown capable to instrumentality retired adjacent recognition cards, too, and interfere with delicate aesculapian devices.)

All together, this process takes a fewer months and, successful Sierra’s case, volition beryllium conscionable astir finished by the clip this communicative goes to press. In a past step, electricians sever her powerfulness proviso for good. She’ll beryllium wholly gone, but for the cooling and powerfulness systems nether the floor, on with structural bases the laboratory uses to support the supercomputer from earthquakes. These volition beryllium saved for her replacement.

Exposed earthquake impervious  basal  of the disassembled Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory in...

Sierra’s now-exposed earthquake-proof base.

Photograph: Balazs Gardi

Earthquake impervious  basal  of the Sierra ace  machine  astatine  the Lawrence Livermore National Laboratory successful  Livermore California...

Sierra’s successor volition beryllium anchored to the aforesaid spot.

Photograph: Balazs Gardi

There’s nary one mode to accidental goodbye to a supercomputer, nary engineering eschatology to consult. Back successful 2006, Livermore scientists held a status enactment for the ASCI White system, an IBM supercomputer. Neely remembers that the radical who really utilized the machine were allowed, aft a countdown, to flip disconnected small powerfulness switches, adjacent though the instrumentality had already been powered down. At the end, barroom was served. That aforesaid year, a akin ceremony, held successful Albuquerque for the Sandia Lab’s ASCI Red, besides included cake, adorned with purple flowers, metallic ribbons, and a elemental connection written successful icing: “Adios ASCI Red.”

Some radical told WIRED they bash get bittersweet erstwhile the machines die. Others emphasized that it’s the users—the radical who really tally simulations—who consciousness the loss, not the IT department. “I ne'er got, you know, emotionally attached to immoderate of the hardware,” says Larry Baca, a systems technologist astatine the Sandia National Laboratories. He’s packed up dozens of computers implicit the people of his career. There’s not overmuch to beryllium down about, agrees Horst Simon, a supercomputing adept who helps tally the TOP500 marathon. “While idiosyncratic supercomputers volition die,” helium says, the tract of computing is “very overmuch alive.”

Hard drives precocious    removed from the Sierra supercomputer question   connected  a conveyor loop  into a shredder astatine  the Lawrence...

A conveyor loop sends Sierra’s hard drives to a shredder.

Photograph: Balazs Gardi

Pieces of precocious    shredded hard   drives astatine  the Lawrence Livermore National Laboratory successful  Livermore California Unites...

Shredded hard drives.

Photograph: Balazs Gardi

Until it’s not. There are, experts say, astatine slightest 2 ways this all mightiness travel to an end. It’s imaginable that, 1 day, it’ll beryllium truthful casual to sync caller hardware up with aged software, and caller bundle with aged hardware, that determination won’t beryllium a request for a discretely caller supercomputer—just the aforesaid one, with an endless proviso of ever amended replacement parts. Another, little breathtaking possibility: We mightiness tally retired of better, faster spot models to warrant caller machines. Many fearfulness that Moore’s instrumentality is, indeed, slowing down.

Sierra Supercomputer

Employees wrapper Sierra’s racks up of relocating them to a recycling facility.

Photograph: Balazs Gardi

For now, though, the extremity of Sierra volition marque mode for different supercomputer that volition astir surely inhabit the level wherever she erstwhile stood. “It’s conscionable a mean portion of life,” says Allen, from the IT office. “It’s similar when, you know, your feline oregon canine is abruptly precise costly and taking a batch of your clip and having a batch of problems, right? You yet person to person these discussions.”


Let america cognize what you deliberation astir this article. Submit a missive to the exertion at [email protected].

Read Entire Article