Slashdot Mirror
Jaguar Supercomputer Being Upgraded To Regain Fastest Cluster Crown
MrSeb writes with an article in Extreme Tech about the Titan supercomputer. From the article: "Cray, AMD, Nvidia, and the Department of Energy have announced that the Oak Ridge National Laboratory's Jaguar supercomputer will soon be upgraded to yet again become the fastest HPC installation in the world. The new, mighty-morphing computer will feature thousands of Cray XK6 blades, each one accommodating up to four 16-core AMD Opteron 6200 (Interlagos) chips and four Nvidia Tesla 20-series GCGPU coprocessors. The Jaguar name will be suitably inflated, too: the new behemoth will be called Titan. The exact specs of Titan haven't been revealed, but the Jaguar supercomputer currently sports 200 cabinets of Cray XT5 blades — and each cabinet, in theory, can be upgraded to hold 24 XK6 blades. That's a total of 4,800 servers, or 38,400 processors in total; 19,200 Opterons 6200s, and 19,200 Tesla GPUs. ... that's 307,200 CPU cores — and with 512 shaders in each Tesla chip that's 9,830,400 compute units. In other words, Titan should be capable of massive parallelism of more than one million concurrent operations. When the server is complete, towards the end of 2012, Titan will be capable of between 10 and 20 petaflops, and should recapture the crown of Fastest Supercomputer in the World from the Japanese 'K' computer."
LLNL will receive their 20 PF machine dubbed Sequoia later this year. IBM's Blue Genes are known for their good ratio of CPU performance/network performance. This allows the MPI codes to scale well. The same is true for vanilla Cray XT5 and XE6 machines, but if upgraded with GPUs then each node receives a significant boost in computational power without increasing the network performance. This leaves the individual nodes bandwidth starved and makes it next to impossible to achieve peak performance in production code. The abysmal ratio of peak performance to actual production performance of China's Tianhe-1A tells the same story.
Maybe they'll achieve peak performance in Linpack, but for everything else Blue Gene/Q will be a much nicer system than Titan. Plus, on Blue Gene you don't have to deal with the heterogeneous system design, which already gave hell to coders on Roadrunner.
BTW: GCGPU should be corrected to GPGPU.
Computer simulation made easy -- LibGeoDecomp
The only thing that concerns me about the name Titan is what happens when they need another name some years from now to indicate its even faster.
Well, Olympian would be the obvious choice for a successor to Titan.
The enemies of Democracy are
The DOE categorizes their supercomputers into capacity machines and capability machines.
The capacity machines are the work horses - time-shared between lots of users doing a variety of applications (including material science, life science, nuclear simulation, etc). The spend pretty much their entire lives near maximum utilization.
The capability machines are the really big ones (Jaguar, Road Runner, etc) that are big enough to permit applications that are too large (require too much RAM or have absurdly long running times) to run on most systems. (Capability machines are also quite difficult to administer because none of the software they run has ever been tested at those scales)
To make laws that man cannot, and will not obey, serves to bring all law into contempt.
--E.C. Stanton
Titan will be a hugely powerful computer. However, fastest supercomputer might be just out of reach. 2012 is also the year that Lawrence Livermore labs, also part of the Department of Energy, is planning to unveil their 20 petaflop BlueGene/Q computer name Seqoia. [http://www.hpcwire.com/hpcwire/2009-02-03/lawrence_livermore_prepares_for_20_petaflop_blue_gene_q.html]
That said, Seqoia will be a classified system for nuclear stockpile simulations. Titan will be a comparatively open system for wide ranging scientific discovery: government, academic, and industrial.
Go Atari!
I would have thought that the world's fastest cluster would something a little more modern than OS X 10.2.
Regaining the #1 spot on the Top500 is merely a convenient side effect. The real reason for building these machines is that they are a key enabler for numerous science projects ranging from astrophysics to climate modeling to atomic phasefield simulation of crystal growth. This type of research can only be done on machines which offer Petabytes of RAM and Petaflops of performance. They cost hundreds of millions to build and operate. And if you can cut this cost down to a fraction by reusing Jaguar's existing housing, cooling and networking facilities, then this is financially a very clever move.
Computer simulation made easy -- LibGeoDecomp
I don't think they'll be using Colossus since the British WWII cryptographers used it first =)
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.