A supercomputing journey inspired by curiosity
History of Cray Supercomputers.
Extending the boundaries of what's possible
It started with the vision of a single individual. Seymour Cray loved the art of designing computers. He loved the challenge of coaxing more and more speed out of circuits and wires. And he loved the allure of doing it all with simplicity and grace.
Mr. Cray formed Cray Research in 1972 to ask new questions. To realize his vision of the world's fastest computing systems. He saw impossible, shrugged, and went back to his work.
With this vision and a small group of engineers, Cray produced the Cray-1 supercomputer. A masterpiece of engineering, the Cray-1 rewrote compute technology from processing to cooling to packaging. And it wrote a company and an industry permanently into history.
From the Cray-1 to HPE Cray
Cray Research's mission at its start was to build the industry's fastest computers and lead large-scale scientific computing. But as the company saw what people could do with those systems, the mission evolved. Simple speed expanded to include a commitment to helping solve scientific and industrial problems that make the world safer, healthier, and smarter.
HPE shared the vision that began with Seymour Cray. In 2019, HPE acquired Cray and with it opened a new chapter in a vibrant, ongoing supercomputing journey.
The 1970s to 1980s: Revolutionary ideas to gigaflops and gallium arsenide
Cray-1 - 1976
First customer: Los Alamos National Laboratory (pictured)
Fun fact: The Cray-1 was the world’s fastest supercomputer from 1976 to 1982. It measured 8½ feet wide by 6½ feet high and contained 60 miles of wires.
Tech story: Cray Research achieved the Cray-1's record-breaking 160 megaflops performance through its small size and cylindrical shape, 1 million-word semiconductor memory, 64-bit word size, optimized Fortran compiler, and a technique called "chaining."
Cray X-MP - 1982
First customer: Digital Productions
Fun fact: Film company Digital Productions used their Cray X-MP system to create images for cult classic The Last Starfighter—one of the first films to use computer-generated imagery.
Tech story: Cray moved into parallel processing with the X-MP system. It featured two shared-memory parallel vector processors with a 9.5 nanosecond clock cycle. Cray produced 11 different versions of the X-MP—the fastest topped out at 941 megaflops.
Cray-2 - 1985
First customer: National Magnetic Fusion Energy Computer Center
Fun fact: The Cray-2 supercomputer earned the nickname "Bubbles." The first ever to use liquid immersion cooling, the system’s circuitry would run so hot that the liquid would boil. So Seymour Cray put it on full display, making the boiling liquid viewable through glass panels.
Tech story: The Cray-2 was a four processor vector architecture with a 256 million 64-bit memory (the largest central memory available on any computer) and 4.1 nanosecond clock speed. It reached a peak speed of 1.9 gigaflops.
Cray Y-MP - 1988
First customer: NASA Ames Research Center
Fun fact: The Cray Y-MP supercomputer has a cameo in the 1992 film Sneakers. The film's leads (Ben Kingsley and Robert Redford) sit on a Y-MP while discussing how to change the world.
Tech story: The Cray Y-MP was the world's first supercomputer to sustain over 1 gigaflops. Considered a follow-on to the X-MP, the initial system had eight parallel vector processors, denser circuits, and a larger central memory. The Y-MP reached a peak speed of 2.67 gigaflops.
The 1990s to 2000s: Change and growth
Cray C90 - 1991
First customer: European Centre for Medium-Range Weather Forecasts
Fun fact: The Ford Motor Company was the first commercial customer for the Cray C90 system.
Tech story: First named the Cray Y-MP C90 supercomputer, the parallel vector system had 16 new CPUs capable of 1 gigaflops each and 2 gigabytes of central memory. It could perform 5 times faster than Cray's previous best. One of its most significant architectural features was a dual-vector pipeline allowing each of the CPUs to deliver two vector results per functional unit every clock period.
Cray Y-MP EL - 1991
Customer fact: Booked orders for 55 systems in its first year alone
Fun fact: The Cray Y-MP EL system was the first supercomputer Cray Research ever sold into the financial services industry and in Eastern Europe.
Tech story: The Y-MP EL supercomputer lowered the cost of entry for customers. It could be configured with up to four processors and 1,042 megabytes of memory. It was air cooled, compact, and easily powered. And with balanced memory and I/O bandwidth, the Y-MP EL offered the highest throughput performance available at its price point. It was later developed into the EL90 series.
Cray M90 - 1992
First customer: AT&T Bell Labs
Fun fact: The Cray M90 supercomputer offered up to 125 times more memory capacity than the original Y-MP system.
Tech story: The Cray M90 supercomputer series (originally called the Y-MP M90) was a large-memory variant of the Cray Y-MP system. Implemented in DRAM technology, it came in two-, four-, or eight-processor configurations with up to 4 billion words of memory. It allowed users to tackle huge problems without breaking them up and running them separately.
Cray T3D - 1993
First customer: Pittsburgh Supercomputing Center
Fun fact: NASA mission planners at the Jet Propulsion Laboratory used a T3D system to explore the best places to land a spacecraft on Mars.
Tech story: The Cray T3D supercomputer was Cray's first massively parallel processing system and the world's first scalable heterogeneous one. It was designed, developed, manufactured, and delivered in just 26 months. The T3D connected a massively parallel array of microprocessors made by Digital Equipment Corporation to a Cray designed vector processor and was available in sizes from 32 to 2,084 processors.
Cray J90 - 1994
Customer fact: Booked 37 advance orders prior to announcement
Fun fact: The J90 series was Cray's fastest-selling product ever with more than 170 systems sold in its first year—40% of them to first-time Cray buyers.
Tech story: The J90 series was a scalable, lower-cost system that came in eight-, 16-, and 32-processor configurations. It achieved 10 to 20 times the bandwidth of competing systems and better price/performance by reducing the complexity of the CPUs and using low-cost and high-speed advanced CMOS (complementary metal oxide semiconductor chips). Prices for the systems started at $225,000.
Cray T90 - 1995
Customer fact: Booked eight advance orders prior to announcement
Fun fact: The T90 system was the world's first wireless supercomputer. Using a pioneering device that connected the processor modules to memory, Cray eliminated all internal wiring. (For comparison, the largest C90 system had 36 miles of it.)
Tech story: The successor to the C90 line, the parallel vector processing Cray T90 series was available with one to 32 processors and provided up to 60 gigaflops of performance. It replaced interconnect wires with electrically activated zero-insertion-force (eZIF) connectors with 400 signal contacts each.
Cray T3E - 1995
First customer: Pittsburgh Supercomputing Center
Fun fact: The Cray T3E was the world's first supercomputer to sustain 1 teraflops on a real-world application. By 1997, T3E systems held 14 of the top 20 spots on the list of the world's top supercomputers.
Tech story: The T3E was unrivaled at efficiently and cost effectively scaling from tens to thousands of processors. What made it different was that every part of the system scaled with the number of processors—interprocessor communication, operating system, I/O, and memory.
Cray SV1 Series - 1998
First customer: Alabama Supercomputer Authority
Fun fact: The Ford Motor Company purchased five Cray SV1 supercomputers, making it the system’s top customer.
Tech story: The SV1 scalable vector system featured processors with a peak performance of four gigaflops, single cabinet nodes with a peak performance up to 32 gigaflops, and system clustering tools that combined to produce a supercomputer capable of 1 teraflops. It introduced innovations such as the world's first vector cache memory. And it beat its competition on price/ performance by 8 times.
Cray X1/X1E - 2002
First customer: U.S. Department of Defense
Fun fact: Cray® X1 processors ran so hot, engineers innovated an evaporative spray technology to keep the modules from overheating.
Tech story: The Cray X1 system fused the processor performance of vector systems with the scalability of massively parallel processing—a feat long-considered impossible. High-performance interconnect and memory subsystems allowed the X1 system to scale to 4,096 processors and a peak performance of 50 teraflops. A major upgrade, the X1E tripled the peak performance and compute density.
Cray XT Series - 2004
First customer: Sandia National Laboratories
Fun fact: Cray broke the petaflops barrier with Oak Ridge National Laboratory's XT5 "Jaguar" system (pictured). Jaguar had 200 cabinets, 224,256 processing cores, and at 1.759 petaflops was the world’s fastest from 2009 to 2010.
Tech story: Cray developed the technology that would become the Cray® XT™ series in partnership with Sandia National Laboratories. The massively parallel XT used a Cray-designed interconnect, was air-cooled, and could scale from a single cabinet to hundreds.
Cray XMT - 2006
First customer: Pacific Northwest National Laboratory
Fun fact: A Georgia Tech researcher used PNNL's Cray® XMT™ system to identify Twitter's most influential users in 2009. The machine chewed through a day's worth of tweets from its ~18 million users in an hour.
Tech story: The Cray XMT system was a scalable, massively multithreaded platform with a shared memory architecture ideally suited for large-scale data analysis and data mining. The design was based on a Cray XT compute blade but used custom Cray Threadstorm chips. A single Threadstorm processor could sustain 128 simultaneous threads.
The 2010s: Harnessing big data
Cray XE6 - 2010
Customer fact: Swiss National Supercomputing Centre (CSCS) received the first beta system.
Fun fact: Scientists using CSCS's Cray® XE6™ system solved a 300-year-old riddle about why Earth's magnetic field keeps moving west. Hint: It’s the core.
Tech story: The XE6 supercomputer took the XT series infrastructure and incorporated multicore AMD processors and the breakthrough Cray Gemini interconnect. Gemini delivered extreme messaging rates and much improved latency. The XE6 could scale to over 1 million processor cores and could exceed 10 petaflops.
Cray XK Series - 2011
First customer: Swiss National Supercomputing Centre
Fun fact: Cray® XK™ system "Titan" at Oak Ridge National Laboratory (pictured) topped the world's fastest list in 2012. It stayed among the top 10 for the next seven years. In 2019, ORNL named a Cray® Shasta™ system "Frontier" as a successor to Titan.
Tech story: The Cray XK series combined the Gemini interconnect, AMD multicore scalar processors, and NVIDIA many-core GPUs to create a true hybrid supercomputer. The system was capable of scaling to 500,000 processors and more than 30 petaflops.
Cray XC Series - 2012
First customer: High Performance Computing Center Stuttgart (pictured)
Fun fact: Rolled out in 2016, the XC50 supercomputer could deliver 1 petaflop of peak performance in a single cabinet. By comparison, the first Cray system to break the petaflops barrier had 200 cabinets.
Tech story: Scalable to 500 petaflops, Cray® XC™ systems merged scalar processing, co-processing, and accelerator technology in a single, highly scalable machine. It also introduced the Aries interconnect which delivered substantial improvements across all metrics and solved the challenge of providing cost-effective, scalable global bandwidth.
Cray Urika Platforms - 2012
Customer fact: The first Cray® Urika® product had several early adopters including the Institute of Systems Biology, Mayo Clinic, Noblis, and Swiss National Supercomputing Centre.
Fun fact: Urika was initially spelled uRiKA. It stood for Universal RDF Integration Knowledge Appliance.
Tech story: The first Urika product was built for a graph-based approach to big data relationship analytics. It used Cray Threadstorm massively multithreaded processors and a massive shared memory architecture. Later called the Urika-GD, it was joined by the Urika-XA extreme analytics and Urika-GX agile analytics platforms.
Cray CS Series - 2013
Customer fact: Railway Technical Research Institute was the first publicly announced customer for the CS series.
Fun fact: What do Shadow, Ollie, BeBop, and Nurion have in common? Hint: They're not pet dogs. They're a few of the CS series systems that have made the TOP500 supercomputer list. Nurion made it all the way to number 11 in 2018.
Tech story: The Cray® CS™ series cluster supercomputers featured the latest in processing, networking, and cooling technologies while also being highly customizable for compute, data, and hybrid workloads. The series' final iteration—the CS500 system—could scale to over 60 petaflops.
Cray CS-Storm - 2014
First customer: U.S. Government
Fun fact: When the Cray® CS-Storm™ system launched in 2014, it was the world's most dense GPU system.
Tech story: Based on the CS series architecture, the CS-Storm supercomputer was built for GPU-accelerated workloads. A single, high-density rack delivered up to 980 teraflops of double-precision performance. The system is well-suited to accelerate computing workloads such as machine and deep learning, reservoir simulation, and geospatial intelligence.
HPE Cray EX - 2018
First customer: Argonne National Laboratory
Fun fact: An exascale HPE Cray supercomputer is 6 billion times faster than the Cray-1.
Tech story: The HPE Cray supercomputer is built for exascale era workloads. It supports a diversity of processor technologies, supports converged workloads, eliminates the distinction between supercomputers and clusters, and fuses HPC and AI workloads with the productivity of the cloud. The revolutionary HPE Slingshot interconnect serves as its backbone. The U.S.'s first three exascale supercomputers are all HPE Cray systems.