The rise and fall of the PlayStation supercomputers
Dozens of PlayStation 3s sit in a refrigerated shipping container on the University of Massachusetts Dartmouth’s campus, sucking up energy and investigating astrophysics. It’s a popular stop for tours trying to sell the school to prospective first-year students and their parents, and it’s one of the few living legacies of a weird science chapter in PlayStation’s history.
Those squat boxes, hulking on entertainment systems or dust-covered in the back of a closet, were once coveted by researchers who used the consoles to build supercomputers. With the racks of machines, the scientists were suddenly capable of contemplating the physics of black holes, processing drone footage, or winning cryptography contests. It only lasted a few years before tech moved on, becoming smaller and more efficient. But for that short moment, some of the most powerful computers in the world could be hacked together with code, wire, and gaming consoles.
Researchers had been messing with the idea of using graphics processors to boost their computing power for years. The idea is that the same power that made it possible to render Shadow of the Colossus’ grim storytelling was also capable of doing massive calculations — if researchers could configure the machines the right way. If they could link them together, suddenly, those consoles or computers started to be far more than the sum of their parts. This was cluster computing, and it wasn’t unique to PlayStations; plenty of researchers were trying to harness computers to work as a team, trying to get them to solve increasingly complicated problems.
The game consoles entered the supercomputing scene in 2002 when Sony released a kit called Linux for the PlayStation 2. “It made it accessible,” Craig Steffen said. “They built the bridges, so that you could write the code, and it would work.” Steffen is now a senior research scientist at the National Center for Supercomputing Applications (NCSA). In 2002, he had just joined the group and started working on a project with the goal of buying a bunch of PS2s and using the Linux kits to hook them (and their Emotion Engine central processing units) together into something resembling a supercomputer.
They hooked up between 60 and 70 PlayStation 2s, wrote some code, and built out a library. “It worked okay, it didn’t work superbly well,” Steffen said. There were technical issues with the memory — two specific bugs that his team had no control over.
“Every time you ran this thing, it would cause the kernel on whatever machine you ran it on to kind of go into this weird unstable state and it would have to be rebooted, which was a bummer,” Steffen said.
They shut the project down relatively quickly and moved on to other questions at the NCSA. Steffen still keeps one of the old PS2s on his desk as a memento of the program.
But that’s not where PlayStation’s supercomputing adventures met their end. The PS3 entered the scene in late 2006 with powerful hardware and an easier way to load Linux onto the devices. Researchers would still need to link the systems together, but suddenly, it was possible for them to imagine linking together all of those devices into something that was a game-changer instead of just a proof-of-concept prototype.
That’s certainly what black hole researcher Gaurav Khanna was imagining over at UMass Dartmouth. “Doing pure period simulation work on black holes doesn’t really typically attract a lot of funding, it’s just because it doesn’t have too much relevance to society,” Khanna said.
Money was tight, and it was getting tighter. So Khanna and his colleagues were brainstorming, trying to think of solutions. One of the people in his department was an avid gamer and mentioned the PS3’s Cell processor, which was made by IBM. A similar kind of chip was being used to build advanced supercomputers. “So we got kind of interested in it, you know, is this something interesting that we could misuse to do science?” Khanna says.
Inspired by the specs of Sony’s new machine, the astrophysicist started buying up PS3s and building his own supercomputer. It took Khanna several months to get the code into shape and months more to clean up his program into a working order. He started with eight, but by the time he was done, he had his own supercomputer, pieced together out of 176 consoles and ready to run his experiments — no jockeying for space or paying other researchers to run his simulations of black holes. Suddenly, he could run complex computer models or win cryptography competitions at a fraction of the cost of a more typical supercomputer.
Around the same time, other researchers were having similar ideas. A group in North Carolina also built a PS3 supercomputer in 2007, and a few years later, at the Air Force Research Laboratory in New York, computer scientist Mark Barnell started working on a similar project called the Condor Cluster.
The timing wasn’t great. Barnell’s team proposed the project in 2009, just as Sony was shifting toward the pared-back PS3 slim, which didn’t have the capability to run Linux, unlike the original PS3. After a hack, Sony even issued a firmware update that pulled OpenOS, the system that allowed people to run Linux, from existing PS3 systems. That made finding useful consoles even harder. The Air Force had to convince Sony to sell it the un-updated PS3s that the company was pulling from shelves, which, at the time, were sitting in a warehouse outside Chicago. It took many meetings, but eventually, the Air Force got what it was looking for, and in 2010, the project had its big debut.
Running on more than 1,700 PS3s that were connected by five miles of wire, the Condor Cluster was huge, dwarfing Khanna’s project, and it used to process images from surveillance drones. During its heyday, it was the 35th fastest supercomputer in the world.
But none of this lasted long. Even while these projects were being built, supercomputers were advancing, becoming more powerful. At the same time, gaming consoles were simplifying, making them less useful to science. The PlayStation 4 outsold both the original PlayStation and the Wii nearing the best-selling status currently held by the PS2. But for researchers, it was nearly useless. Like the slimmer version of the PlayStation 3 released before it, the PS4 can’t easily be turned into a cog for a supercomputing machine. “There’s nothing novel about the PlayStation 4, it’s just a regular old PC,” Khanna says. “We weren’t really motivated to do anything with the PlayStation 4.”
The era of the PlayStation supercomputer was over.
The one at UMass Dartmouth is still working, humming with life in that refrigerated shipping container on campus. The UMass Dartmouth machine is smaller than it used to be at its peak power of about 400 PlayStation 3s. Parts of it have been cut out and repurposed. Some are still working together in smaller supercomputers at other schools; others have broken down or been lost to time. Khanna has since moved on to trying to link smaller, more efficient devices together into his next-generation supercomputer. He says the Nvidia Shield devices he’s working with now are about 50 times more efficient than the already-efficient PS3.
It’s the Air Force’s supercluster of super consoles that had the most star-studded afterlife. When the program ended about four years ago, some consoles were donated to other programs, including Khanna’s. But many of the old consoles were sold off as old inventory, and a few hundred were snapped up by people working with the TV show Person of Interest. In a ripped-from-the-headlines move, the consoles made their silver screen debut in the show’s season 5 premiere, playing — wait for it — a supercomputer made of PlayStation 3s.
“It’s all Hollywood,” Barnell said of the script, “but the hardware is actually our equipment.”
Correction, 7:05 PM ET: Supercomputer projects needed the original PS3, not the PS3 Slim, because Sony had removed Linux support from the console in response to hacks — which later led to a class-action settlement. This article originally stated that it was because the PS3 Slim was less powerful. We regret the error.