IBM Designed Military Supercomputer Sets New Record

[malware]

A government computer in New Mexico is the first supercomputer to perform at one petaflop (one thousand trillion calculations per second). Located at the Los Alamos National Laboratory of the Department of Energy's National Nuclear Security Administration, Roadrunner is twice as fast as IBM Blue Gene system at Lawrence Livermore National Lab, which was until now the fastest computer in the world. The new supercomputer is designed and built by IBM using both traditional computer chips and IBM's Cell Broadband Engine. Roadrunner occupies 6,000 square feet and weighs 500,000 lbs. It is also aiming to take place among the top energy-efficient systems on the official "Green 500" list of supercomputers. Roadrunner will be used primarily to ensure the safety and reliability of the U.S. nuclear weapons stockpile. It will also do research into astronomy, energy, human genome science and climate change. Learn more here.



View at TechPowerUp Main Site

 

[Solaris17]

sweet.

 

[pentastar111]

Keeping tabs on nuclear materials is fine and dandy...but how are the FPS when playing Crysis?? Just kidding , just kidding!

 

[Cold Storm]

That guy looks like my old AM History teacher! woot woot! I wanna take a tour and see that with my own eyes!

 

[DarkMatter]

I wonder how many FASTRA you would need to mach that computer. A lot less than what we could think at first, surely.

 

[PVTCaboose1337]

I love looking at those rooms... they have aisles and everything. Cute.

 

[jbunch07]

sweet!
how much does it cost?

 

[WarEagleAU]

Reminds me of back in the 60s and 70s when one computer (a desktop we call it today) took up as much if not more room than this super computer and weighed as much if not more. Anyone remember War Games with Matthew Broderick and Ally Sheedy? W.O.P.R.

 

[imperialreign]

when I read the title (IBM Designed Military Supercomputer Sets New Record), I thought the record was going to be a new average FPS in Crysis with everything maxed



Still impressive, though

 

[lemonadesoda]

• Roadrunner, named after the New Mexico state bird, cost about $100 million
• Roadrunner is the world’s first hybrid supercomputer. In a first-of-a-kind design, the Cell Broadband Engine® -- originally designed for video game platforms such as the Sony Playstation 3® -- will work in conjunction with x86 processors from AMD®.
• Made from Commercial Parts. In total, Roadrunner connects 6,948 dual-core AMD Opteron® chips (on IBM Model LS21 blade servers) as well as 12,960 Cell engines (on IBM Model QS22 blade servers). The Roadrunner system has 80 terabytes of memory, and is housed in 288 refrigerator-sized, IBM BladeCenter® racks occupying 6,000 square feet. Its 10,000 connections – both Infiniband and Gigabit Ethernet -- require 57 miles of fiber optic cable. Roadrunner weighs 500,000 lbs. Companies that contributed components and technology include; Emcore, Flextronics, Mellanox and Voltaire.
• Custom Configuration. Two IBM QS22 blade servers and one IBM LS21 blade server are combined into a specialized “tri-blade” configuration for Roadrunner. The machine is composed of a total of 3,456 tri-blades built in IBM’s Rochester, Minn. plant. Standard processing (e.g., file system I/O) is handled by the Opteron processors. Mathematically and CPU-intensive elements are directed to the Cell processors. Each tri-blade unit can run at 400 billion operations per second (400 Gigaflops).
• Roadrunner operates on open-source Linux software from Red Hat.

 

[DarkMatter]

Reminds me of back in the 60s and 70s when one computer (a desktop we call it today) took up as much if not more room than this super computer and weighed as much if not more. Anyone remember War Games with Matthew Broderick and Ally Sheedy? W.O.P.R.

Hmm dude, those computers were not what we call today desktop PCs... They were the supercomputers of that time, which happened to perform like desktops we have today.

 

[overclocker!]

looks like my new pc

 

[lemonadesoda]

Remember that these new-style "supercomputers" are actually a shed of small PCs. They cannot approach anywhere near one petaflop (one thousand trillion calculations per second) on a single threaded calculation algorithm. What is also important is whether this statistic is theoretic CPU core flops or actual software delivered flops.

What happens is that the network needs to be managed with a scheduler and marshaller to send "task" over ethernet to each blade. So if you can partition a problem so that separate blades can do different INDEPENDENT calculations, then great.

In other words, its really a fancy "folding at home" but in one room with better interconnect. F@H is currently achieving 2.0 Petaflops at ACTUAL software delivered performance, over 300,000 active CPUs. http://fah-web.stanford.edu/cgi-bin/main.py?qtype=osstats

The F@H network is more actual performance than Roadrunner in terms of theoretic CPU flops. That's pretty cool for F@H.

 

[Kei]

We must kill skynet...btw D@MN that is fast!

K

 

[DarkMatter]

Remember that these new-style "supercomputers" are actually a shed of small PCs. They cannot approach anywhere near one petaflop (one thousand trillion calculations per second) on a single threaded calculation algorithm. What is also important is whether this statistic is theoretic CPU core flops or actual software delivered flops.

What happens is that the network needs to be managed with a scheduler and marshaller to send "task" over ethernet to each blade. So if you can partition a problem so that separate blades can do different INDEPENDENT calculations, then great.

In other words, its really a fancy "folding at home" but in one room with better interconnect. F@H is currently achieving 2.0 Petaflops at ACTUAL software delivered performance, over 300,000 active CPUs. http://fah-web.stanford.edu/cgi-bin/...?qtype=osstats

The F@H network is more actual performance than Roadrunner in terms of theoretic CPU flops. That's pretty cool for F@H.

F@H does ONLY 0.2 petaflops? I thought it was a lot more than that. They said they could achieve 100 more power with F@H than with any supercomputer. Looking at some supercomputer statistics it seems that the ACTUAL performance is around 60-80% of the peak performance, so this computer is fairly powerful. Now I know they can't use the whole supercomputer for them or 24/7 but I thought F@H was just more.

EDIT: Oh and BTW. The only difference between these "new-style supercomputers" and the "good-old supercomputers" is that the new ones use "standard" open net protocol such as Ethernet instead of the propietary-made-only-for-that-computer ones used in previos supercomputers. The working mechanism is the same.

 

[Conti027]

OC it!!! rofl

 

[warup89]

this makes me remember of the ENIAC, pretty soon that PC will be sitting on an office desk.

 

[Sapientwolf]

Yeah Los Alamos National Laboratories and Sandia National Laboratories always have a lot of cool stuff. I live in Albuquerque, New Mexico, so I usually hear a lot about this kind of thing on the local news.

Good stuff, New Mexico is a great place for high tech research and industry, the more the better.

 

[panchoman]

sweett and DAMN they put opertons in that thing!

and i was wondering when ibm would put together a cell based supercomputer (which would obviously be faster then the ppc-based bluegenes)

looks like this is part of the amd-ibm deal a while back

 

[lemonadesoda]

EDIT: Oh and BTW. The only difference between these "new-style supercomputers" and the "good-old supercomputers" is that the new ones use "standard" open net protocol such as Ethernet instead of the propietary-made-only-for-that-computer ones used in previos supercomputers. The working mechanism is the same.

Not exactly. The "old-style" were vector based architectures, like the Cray's of yesteryear. They didnt suffer from the von Neumann control bottleneck when scaling like modern clusters, esp. Beowulf that is now very common. However the Crays were ultra powerful at vector-scalable problems but easily outperformed by much cheaper scalar CPUs for more regular (and easier to program) computing tasks.

Pricing for COTS has just made Cluster supercomputing so cheap relative to SIMD and Vector that SIMD and Vector is essentially dead. However, if you have a SIMD or vector type problem and try to cluster it you hit von Neumann very fast where the marginal gain from an extra node has exponentially decreasing returns.

Anyone want to build their own supercomputer?

http://www.ibiblio.org/pub/linux/docs/HOWTO/archive/Beowulf-HOWTO.html
http://www.beowulf.org/overview/index.html

 

[mlee49]

Someone hack it to F@home for TPU! We could easily climb number one within an hour!
EDIT 1 minute or less

 

[panchoman]

Someone hack it to F@home for TPU! We could easily climb number one within an hour!

hows about less then a minute lol

 

[lemonadesoda]

Great idea. W1z should program a nifty tool called TPUpi which connects all logged in TPU members, to run a networked version of superPI. I bet our community FLOPS would pwn. :grin:

 

[trt740]

I just bought one on ebay

 

[jbunch07]

I just bought one on ebay

how much you pay for it?