Tikalon Header

Aurora Exascale Computer

June 10, 2019

During the first decades of electronic computers that began in earnest with the construction of ENIAC in 1945, all computing was done on room-filling mainframe computers. While these proved useful for tasks such as printing paychecks and facilitating the census, scientists and engineers needed faster computers. Thus arose a different kind of computer called the supercomputer.

An important milestone in supercomputer design was the Cray-1, named for its talented designer, Seymour Cray (1925-1996), who is known as the father of supercomputing. The Cray supercomputer's innovation was in its implementation of vector processing, which was a huge advance over the scalar processing motif of the mainframes. The first Cray-1 was delivered to Los Alamos National Laboratory in 1976.

Integrated circuit technology was not very advanced at the time of the Cray's introduction, There were not many transistors per chip, the transistors were bipolar, not field effect, and high-speed circuitry generated quite a bit of heat. For that reason, liquid cooling using Freon was required for its power-hungry ECL logic chips. The Cray-1 floating point performance was about 80 million floating-point operations per second (FLOP).[1] The supercomputers of today are in the hundreds of petaflop range, which is a million times faster.

Cray-1 supercomputer at the Wisconsin Historical Museum, Madison, Wisconsin.


Cray-1 supercomputer at the Wisconsin Historical Museum, Madison, Wisconsin.

As an illustration of how technology has progressed, a typical smartphone is about ten times faster and nearly 30,000 times lighter than the Cary-1.

(Wikimedia Commons image by Daderot.)


The above photograph shows the unusual shape of the Cray-1. This shape wasn't dictated by aesthetics alone. The speed of light is such that it takes 3.3 nsec for light to travel one meter. Electrical signals in wire travel about a meter in 5 nsec. Converting time to frequency shows that electrical signals can only traverse a meter at a rate of 200 MHz. The clock rate of the Cary-1 was 80 MHz.

At about the same time that the Cray-1 was introduced, the pre-Internet meme for a more powerful supercomputer was the "hairy, smoking, golf ball." Golf ball indicates the small size needed because of the limitation imposed by the speed of light, while hairy indicates all the wires needed to exchange data with the outside world, and a 128-bit data word would require 128 wires plus a signal return. Indeed, the ball grid array interconnects of today's complex integrated circuits are equivalent of having nearly 500 wires. This computer would be smoking, since the power dissipation of the silicon circuitry used at that time was about two watts per square inch, and a lot of silicon would be needed inside the golf ball.

Integrated circuit technology has improved over the decades, so modern supercomputers never entered a hairy, smoking golf ball phase. The earliest useful microprocessor, the Intel 8080, had 4,500 transistors based on 6,000 nm lithography. Today's 32-core AMD Epyc processor has nearly 20 billion transistors that are rendered by a 14 nm lithographic process onto a 768 mm2 die.

The Summit supercomputer at Oak Ridge National Laboratory.

The Summit supercomputer at Oak Ridge National Laboratory (ORNL, Oak Ridge, Tennessee). This supercomputer presently holds first place on the top500.org list of supercomputers. (Oak Ridge National Laboratory image)


There's a website, top500.org, that keeps track of the world's fastest computers. Here's a list of the top five as of this writing:

#1 Summit, a supercomputer at Oak Ridge National Laboratory (ORNL, Oak Ridge, Tennessee), has a performance of 143.5 petaflop/second on a software benchmark written in the HPL language. Summit is built from 4,356 computing nodes incorporating GPUs communicating via InfiniBand.

#2 Sierra, a supercomputer at Lawrence Livermore National Laboratory (Livermore, California) has an architecture very similar to Summit. It's built with 4,320 nodes to achieve 94.6 petaflop/second.

#3 Sunway TaihuLight is a supercomputer developed by China's National Research Center of Parallel Computer Engineering & Technology (NRCPC, Wuxi, China). It has a performance of 93 petaflop/second, quite close to that of #2 Sierra.

#4 Tianhe-2A (Milky Way-2A) is a supercomputer developed by China's National University of Defense Technology (NUDT, Changsha, China) with a performance of 61.4 petaflop/second.

#5 Piz Daint, a Cray XC50 system with 361,760-cores installed at the Swiss National Supercomputing Centre (CSCS, Lugano, Switzerland) has a performance of 21.2 petaflop/second. Note the large gap between this supercomputer and the next highest on the list.

We're at the point in computational performance at which the metric prefix beyond peta- (1015) will be breached. The next step is exa- (1018) and exascale computing with 1018 floating point operations per second. The push towards exascale computing in the United States began in July, 2015, with the creation of a National Strategic Computing Initiative. In March, 2019, the United States Department of Energy announced that the first exaFLOP supercomputer, called Aurora, a joint development of Argonne National Laboratory (Argonne, Illinois), Intel and Cray, would be operational at Argonne by the end of 2021.[2-4]

Aurora, which will be sited at the Argonne Leadership Computing Facility that was established in 2006, will be constructed at a cost of more than $500 million.[2-4] Says United States Secretary of Energy, Rick Perry,
"Achieving exascale is imperative, not only to better the scientific community, but also to better the lives of everyday Americans... Aurora and the next generation of exascale supercomputers will apply HPC and AI technologies to areas such as cancer research, climate modeling and veterans' health treatments. The innovative advancements that will be made with exascale will have an incredibly significant impact on our society."[2]

Aurora will include several technological innovations, such as a new I/O system, the Distributed Asynchronous Object Store (DAOS), a future generation of Intel's Xeon architecture, and Cray's next-generation Shasta platform contained in 200 cabinets. Aurora is designed specifically to open high-performance computing to artificial intelligence applications.[2,4] It's expected that Aurora will be used to investigate questions in physical cosmology, alternative energy sources, the design safer vehicles, the invention of new materials, and neuroscience.[3]

Aurora Supercomputer

The proposed Aurora supercomputer that will be sited at the Argonne Leadership Computing Facility of Argonne National Laboratory. (Argonne National Laboratory image.)


To these ends, the DOE has established the Aurora Early Science Program of 15 projects whose project teams will collaborate with computer scientists in code migration and optimization for the Aurora.[3] Some of these projects are listed below:[3]
· Extreme-Scale Cosmological Hydrodynamics, Katrin Heitmann, Principal Investigator.

· Exascale Computational Catalysis, David Bross, Principal Investigator.

· Dark Sky Mining, Salman Habib, Principal Investigator.

· Enabling Connectomics at Exascale to Facilitate Discoveries in Neuroscience, Nicola Ferrier, Principal Investigator.

References:

  1. Neal Pritchett, "Mini Cray," notpurfect.com.
  2. U.S. Department of Energy and Intel to deliver first exascale supercomputer, Argonne National Laboratory Press Release, March 18, 2019.
  3. Aurora - Coming in 2021, Argonne National Laboratory Website.
  4. Stephen Mraz, "Intel and Cray Will Deliver First Exoscale Supercomputer to National Lab, Electronic Design, April 24, 2019.
  5. Aurora: Argonne’s Next-Generation Exascale Supercomputer, YouTube Video by Argonne National Laboratory, March 19, 2019.
  6. Cray Announces First Exascale System, YouTube Video by Cray Inc., March 18, 2019.