Tikalon Blog by Dev Gualtieri

Titan Supercomputer

November 19, 2012

In a previous article (Special K, June 23, 2011), I wrote about the fastest computer at that time, the K Computer of the RIKEN Advanced Institute for Computational Science , Kobe, Japan . This computer was benchmarked at 8.2 petaflop , and the ten fastest computers at that time all achieved petaflop performance.[1] Compare that with the 80 megaflop performance of one of the first supercomputers , the venerable Cray-1 .

There will always be another, faster computer, and there's an organization dedicated to cataloging supercomputing progress. Supercomputer rankings are posted every six months in the TOP500 List of the fastest supercomputers.[2] The ranking is determined by the computers' speed of execution of the Linpack benchmark suite. I wrote about Linpack in a previous article (Benchmarks, November 17, 2010).

November's release of its 40th list marked the Top 500's 20th anniversary. Here's a summary of the current top ten.[2] The list shows that maximum computing speed has jumped from 8.2 to 17.6 petaflops in just eighteen months. It's interesting to note that the second ranked computer has more cores and a higher clock rate than the first, which proves that system architecture is as important as the hardware .

Rank	Country	Site	System	Cores	F(GHz)	Pflop
1	United States	DOE /ORNL	Titan - Cray XK7 , Opteron 6274 16C	560640	2.20	17.6
2	United States	DOE/LLNL	Sequoia - BlueGene/Q, Power BQC 16C	1572864	2.60	16.3
3	Japan	RIKEN	K Computer - SPARC64 VIIIfx	705024	2.00	10.5
4	United States	DOE/Argonne National Laboratory	Mira - BlueGene/Q, Power BQC 16C	786432	1.60	8.2
5	Germany	Forschungs-zentrum Juelich	JUQUEEN - BlueGene/Q, Power BQC 16C	393216	1.60	4.1
6	Germany	Leibniz Rechenzentrum	SuperMUC - iDataPlex DX360M4, Xeon E5-2680 8C	147456	2.70	2.9
7	United States	University of Texas	Stampede - PowerEdge C8220, Xeon E5-2680 8C	204900	2.70	2.7
8	China	Tianjin National Supercomputing Center	Tianhe-1A - NUDT YH MPP, Xeon X5670 6C	186368	2.93	2.6
9	Italy	CINECA	Fermi - BlueGene/Q, Power BQC 16C	163840	1.60	1.7
10	United States	IBM Development Engineering	DARPA Trial Subset - Power 775, POWER7 8C	63360	3.84	1.5

As seen in the list, the Oak Ridge National Laboratory Titan computer is the fastest in the world (except, perhaps, for some NSA computers about which we have no information).[3-10] One important note is the power required for such speed. The Titan supercomputer draws 8.2 megawatts , which is surprisingly lower than the K computer's 12.66 megawatts. When you model climate change with one of these, you're changing the climate at the same time! A megawatt translates to hundreds of dollars per hour operating cost.

Computation sometimes leads to a headache, as the leftmost figure demonstrates.

The Titans , as illustrated by Gustave Doré (1832 - 1883) in Dante's Inferno (Plate LXV: Canto XXXI).

(Via Wikimedia Commons).

As described in an ORNL press release, the computing power of the Titan is equivalent to each of the Earth's seven billion inhabitants doing three million calculations per second.[4] These aren't simple calculations, but ones that would put so-called long division to shame. Humans, however, should be proud that we're still the only intelligent lifeforms on the planet, at least for a few more years (My apologies to the dog owners, dolphin lovers, etc., who may object to my saying that humans are Earth's only intelligent lifeform).

There's a trend towards the use of Graphics processing unit (GPU) chips in computing applications. There are sixty-two supercomputers on the current top-500 list using GPUs to increase speed and efficiency . In the top-10, these include the Titan and the Chinese Tianhe-1A. Both systems use Nvidia GPUs.[3] Says ORNL associate laboratory director for computing and computational sciences, Jeffrey A. Nichols ,

"One challenge in supercomputers today is power consumption... Combining GPUs and CPUs in a single system requires less power than CPUs alone and is a responsible move toward lowering our carbon footprint."[4]

That's one hot supercomputer.

The ORNL Titan supercomputer consumes 8.2 megawatts of power.

(ORNL Image).

Titan is actually an upgrade of an ORNL supercomputer called Jaguar , the upgrade being essentially a replacement of its six-core AMD Opteron CPUs to 16-core, plus an integration of Nvidia Tesla K20 GPUs into the system to increase computing performance by a factor of ten.[4-5] A GPU was added to each of the 18,688 AMD Opteron CPU chips.[7] Titan contains more than 700 terabytes of memory,and sits on 20,000 square feet of floor area. There's enough brute-force cooling that ear protection is required if working more than fifteen minutes in Titan's proximity.[4,7]

Supercomputing won't stop at the 10-petaflop stage. The US Department of Energy is planning an exascale system , one that performs at 10¹⁸ flops, for 2018.[5] ORNL's Jeffrey Nichols, as quoted at datacenterknowledge.com, explains,

"We envision two systems beyond Titan to achieve exascale performance by about 2018... The first will be an order of magnitude more powerful than Titan, in the range of 200 petaflops. This system will be an exascale prototype, incorporating many of the hardware approaches that will be incorporated at the exascale. We hope to scale this solution up to the exascale."[5]

Power-wise, exascale won't be that easy to reach. Wu Feng , who curates the Green-500 list of the most energy-efficient supercomputers, says that an exascale computer will require 50 megawatts if current efficiency trends continue.[7] The DOE's goal is 20 megawatts, or less, so new chip types will be required. Also, interconnection between subsystems would need to be fibreoptic .[7]

The US now hosts half of the computers on the Top-500, with Europe having 105, and Asia 124.[3] The BBC states that supercomputing has joined theory and experiment to become the "third pillar" of doing science .[7]

Titan Supercomputer

References: