'supercomputers' Episodes

Cray Supercomputers


Today we’re going to talk through the history of Cray Computers.

And really, this is then a history of supercomputers during Seymour Cray’s life. If it’s not obvious by his name, he was the founder of Cray. But before we go there, let’s back up a bit and talk about some things that were classified for a long time. The post-World War II spending by the US government definitely leveled up the US computer industry. And defense was the name of the game in those early years. 


Once upon a time, the computer science community referred to the Minneapolis/St Paul area as the Land of 10,000 Top Secret Projects. And a lot of things ended up coming out of that. One of the most important in the history of computing though, was Engineering Research Associates, or ERA. They built highly specialized computers. Those made for breaking Soviet codes. 


Honeywell had been founded in Minneapolis and as with Vannevar Bush, had gone from thermostats to computers. Honeywell started pumping out the DATAmatic 1000 in 1957. There was a computer shipping and Honeywell was well situated to capitalize on the growing mainframe computer market. 


ERA had some problems because the owners were embroiled in Washington politics and so they were acquired by Sperry Rand, today’s Unisys, but at the time one of the larger mainframe developers and the progeny of both the Harvard Mark series and ENIAC series of mainframes. Only problem is that the Sperry Rand crew were making a bundle off Univacs and so didn’t put money into forward looking projects. 


The engineers knew that there were big changes coming in computing. And they wanted to be at the forefront. Who wouldn’t. But with Sperry Rand barely keeping up with orders they couldn’t focus on R&D the way many former ERA engineers wanted to. So many of the best and brightest minds from ERA founded Control Data Corporation, or CDC. And CDC built some serious computers that competed with everyone at the time. Because they had some seriously talented engineers. One, who had come over from ERA, was Seymour Cray. And he was a true visionary. 


And so you had IBM and their seven biggest competitors, known as Snow White and the Seven Dwarfs. Three of those dwarfs were doing a lot of R&D in Minneapolis (or at least the Minneapolis area). None are still based in the Twin Cities. But all three build ruggedized computers that could withstand nuclear blasts, corrosive elements, and anything you could throw at them. 


But old Seymour. He wanted to do something great. Cray had a vision of building the fastest computer in the world. And as luck would have it, transistors were getting cheaper by the day. They had initially been designed to use germanium but Seymour Cray worked to repackage those at CDC to be silicon and was able to pack enough in to make the CDC 6600 the fastest computer in the world in 1964. They had leapfrogged the industry and went to market, selling the machines like hotcakes. 


Now CDC would build one of the first real supercomputers in that 6600. And supercomputers are what Cray is known for today. But there’s a little more drama to get from CDC to Cray and then honestly from Cray to the other Crays that Seymour founded. CDC went into a big of a buying tornado as well. As with the Univacs, they couldn’t keep up with demand and so suddenly were focused too much on Development to look beyond fulfillment and shipping and into the Research part of R&D. Additionally shipping all those computers and competing with IBM was rough and CDC was having financial problems, so CEO William Norris wouldn’t let them redesign the 6600 from the ground up. 


But Cray saw massive parallel processing as the future, which is kinda’ what supercomputing really is at the end of the day, and was bitten by that bug. He wanted to keep building the fastest computers in the world. And he would get his wish. He finally left CDC in 1972 and founded Cray Research along with cofounding engineer Lester Davis. They went to Chippewa Falls Wisconsin.


It took him four years, but Cray shipped the Cray-1 in 1976, which became the best selling supercomputer in history (which means they sold more than 80 and less than a hundred). It was 80MhZ, or 200 gigaFLOPS. And that was vector processing. They would math faster by re-arranging the memory and registers to more intelligently process big amounts of data. He used Maxwell’s equations on his boards. He designed it all on paper. The first Cray-1 would ship to Los Alamos National Laboratory. The Cray-1 was 5 and a half tons, cost around $8 million dollars in 1976 money and the fact that they were the fastest computer in the world combined with the fact that they were space age looking gave Seymour Cray instant star status. 


The Cray-1 would soon get competition from the ILLIAC IV out of the University of Illinois, an ARPA project. So Cray got to work thinkin’. He liked to dig when he thought, and he tried to dig a tunnel under his house. This kinda’ sums up what I think of Wisconsin. 


The Cray-2 would come in 1985, which was the first multiple CPU design by Cray. It came in at 1.9 Gigaflops. They rearranged memory to allow for more parallelization and used two sets of memory registers. It effectively set the stage for modern processing architectures in a lot of ways, offloading tasks for a dedicated foreground processor to main memory connected over the fastest channels possible to each CPU. But IBM wouldn’t release the first real multi core processor until 2001. And we see this with supercomputers. The techniques used in them come downmarket over time. 


But some of the biggest problems were how to keep the wires close together. The soldering of connecters at that level was nearly impossible. And the thing was hot. So they added, get this, liquid coolant, leading some people to call the Cray-2 “Bubbles.”


By now, Seymour Cray had let other people run the company and thee were competing projects like the Cray X-MP underway. Almost immediately after the release of the Cray-2 Seymour moved to working on the Cray-3 but the project was abandoned and again, Cray found himself wanting to just go do research without priorities shifting what he could do. 


But Seymour always knew best. Again, he’s from Wisconsin. So he left the company with his name and started another company, this one called Cray Computer, where he did manage to finish the Cray-3. But that Cold War war spending from the Cold War dried up. And while he thought of designs for a Cray-4, the company would go bankrupt in 1995. He was athletic and healthy, so in his 70s, why not keep at it?


His next company would focus on massively parallel processing, which would be the trend of the future, but Seymour Cray died from complications to a car accident in 1996. 


He was one of the great pioneers of the computing industry. He set a standard that computers like IBM’s Blue Gene then Summit or China’s Sunway TahuLight or Dell’s Frontera or Cray’s HPE or Fujitsu’s aBCI or Lenovo’s SuperMUC-NG carry on. Those run at between 20 gigaflops to close to 150 gigaflops. Today, the Cray X1E pays homage to it’s ancestor, the great Cray-1. 


But no one does it with style the way the Cray-1 did - and think about this, Moore’s Law says transistors will double every two years. Not to oversimplify things but that means that since the Cray-2 we should have had a 262 gigaflop machine by now. But I guess he’s not here to break down the newer barriers like he did with the von Neumann bottleneck. 


Also, think about this, those early supercomputers were funded by the departments that became the NSA. They even helped fund the development of Cray’s throughout history. So maybe we have hit 262 and it’s just classified. I swoon at that thought. But maybe it’s just that this is where the move from bits to qubits and quantum computing becomes the next significant jump. Who knows?


But hey, thanks for joining me on this episode of the History of Computing Podcast. Do you have a story you want to tell? I plan to run more interviews soon and while we have a cast of innovators that we’re talking to, we’d love even more weird and amazing humans. Hit us up if you want to! And in the meantime, thanks again for listening, we are so lucky to have you. 




All About Amdahl


Gene Amdahl grew up in South Dakota and as with many during the early days of computing went into the Navy during World War II. He got his degree from South Dakota State in 1948 and went on to the University of Wisconsin-Madison for his PhD, where he got the bug for computers in 1952, joining the ranks of IBM that year.

At IBM he worked on the iconic 704 and then the 7030 but found it too bureaucratic. And yet he came back to become the Chief Architect of the IBM S/360 project. They pushed the boundaries of what was possible with transistorized computing and along the way, Amdahl gave us Amdahl’s Law, which is an important aspect of parallel computing - how much latency tasks take when split across different CPUs. Think of it like the law of diminishing returns applied to processing. Contrast this with Fred Brook’s Brook’s Law - which says that adding incremental engineers don’t make projects happen faster by the same increment, or that it can cause a project to take even more time.

As with Seymour Cray, Amdahl had ideas for supercomputers and left IBM again in 1970 when they didn’t want to pursue them - ironically just a few years after Thomas Watson Jr admitted that just 34 people at CDC had kicked IBM out of their leadership position in the market.

First he needed to be able to build a computer, then move into supercomputers. Fully transistorized computing had somewhat cleared the playing field. So he developed the Amdahl 470V/6 - more reliable, more pluggable, and so cheaper than the IBM S/370. He also used virtual machine technology so customers could simulate a 370 and so run existing workloads cheaper. The first went to NASA and the second to the University of Michigan. During the rise of transistorized computing they just kept selling more and more machines.

The company grew fast, taking nearly a quart of the market share. As we saw in the CDC episode, the IBM antitrust case was again giving a boon to other companies. Amdahl was able to leverage the fact that IBM software was getting unbundled with the hardware as a big growth hack. As with Cray at the time, Amdahl wanted to keep to one CPU per workload and developed chips and electronics with Fujitsu to enable doing so.

By the end of the 70s they had grown to 6,000 employees on the back of a billion dollars in sales. And having built a bureaucratic organization like the one he just left, he left his namesake company much as Seymour Cray had left CDC after helping build it (and would later leave Cray to start yet another Cray). That would be Trilogy systems, which failed shortly after an IPO. I guess we can’t always bet on the name. Then Andor International. Then Commercial Data Servers, now a part of Xbridge systems.

Meanwhile the 1980s weren’t kind to the company with his name on the masthead. The rise of Unix and first minicomputers then standard servers meant people were building all kinds of new devices. Amdahl started selling servers, given the new smaller and pluggable form factors. They sold storage. They sold software to make software, like IDEs. The rapid proliferation of networking and open standards let them sell networking products.

Fujitsu ended up growing faster and when Gene Amdahl was gone, in the face of mounting competition with IBM, Amdahl tried to merge with Storage Technology Corporation, or StorageTek as it might be considered today. CDC had pushed some of its technology to StorageTek during their demise and StorageTek in the face of this new competition ended up filing Chapter 11 and getting picked up by Sun for just over $4 billion.

But Amdahl was hemorrhaging money as we moved into the 90s. They sold off half the shares to Fujitsu, laid off over a third of their now 10,000 plus workforce, and by the year 2000 had been lapped by IBM on the high end market. They sold off their software division, and Fujitsu acquired the rest of the shares. Many of the customers then moved to the then-new IBM Z series servers that were coming out with 64 bit G3 and G4 chips. As opposed to the 31-bit chips Amdahl, now Fujitsu under the GlobalServer mainframe brand, sells.

Amdahl came out of the blue, or Big Blue. On the back of Gene Amdahl’s name and a good strategy to attack that S/360 market, they took 8% of the mainframe market from IBM at one point. But they sold to big customers and eventually disappeared as the market shifted to smaller machines and a more standardized lineup of chips. They were able to last for awhile on the revenues they’d put together but ultimately without someone at the top with a vision for the future of the industry, they just couldn’t make it as a standalone company.

The High Performance Computing server revenues steadily continue to rise at Fujitsu though - hitting $1.3 billion in 2020. In fact, in a sign of the times, the 20 million Euro PRIMEHPC FX700 that’s going to the Minho Advanced Computing Centre in Portugal is a petascale computer built on an ARM plus x86 architecture. My how the times have changed. But as components get smaller, more precise, faster, and more mass producible we see the same types of issues with companies being too large to pivot quickly from the PC to the post-PC era. Although at this point, it’s doubtful they’ll have a generations worth of runway from a patron like Fujitsu to be able to continue in business. Or maybe a patron who sees the benefits downmarket from the new technology that emerges from projects like this and takes on what amounts to nation-building to pivot a company like that. Only time will tell.

(OldComputerPods) ©Sean Haas, 2020