Damn...Fastest Supercomputer (NASA) uses Intel

probably avoiding their hardware being exposed to missing/out of commission rovers and poorly running shuttles?

 

I don't know what you mean by saying 'poorly running shuttles'. NASA has done a damn good job in my opinion.


As for the super computer, good for NASA.... and Intel.

 

I agree with 'poorly running shuttles'. NASA's inability to keep up technology and get funding from the government is starting show its backside. When was the last time man was on the moon? Its about time they did something more. If its not about racing the Kremlin, do it for mankind. Laziness.

1. Leave the Hubble alone
2. Fix the ISS. Its faulty.
3. Whatever happened to the Mars expedition? Are we just going to sweep sand off Mars?
4. Ugrade to the Space Shuttles. They're inefficient and dangerous. Design new Shuttles. The amount of fuel that needs to be used is astronomical.
5. Find out what really happened to the Columbia.

Just a few thoughts.

 

Attack of the eco-weenies, the freon in the spray-on insulating foam had to be removed from the mix to make all the eco-weenies happy.

Since they removed the freon, they've been having problems with the foam on the fuel tank.

Since they could exactly shut the program down while they spend a year fixing the problem/formula, they motored on through it -- and labeled the foam/ice falling off the tank an acceptable risk -- even when it start coming off in big chunks.

 

Apple don't make large-scale multiprocessing systems and SGI do. Each of these systems has 512 CPUs, and the most you can get in a Mac is 2.

Teraflops can make for interesting numbers, but not all problems lend themselves well to simple clusters, which is all that Apple can offer.

 

Something I find interesting: I have seen a couple of news articles on this story already, and neither one mentions how much it cost to build this system. SGI may not be touting the price because they probably still don't compete with Apple on a $/teraflops basis.

I could be wrong, but I seem to recall that the original VT supercomputer project (2GHz G5 units) cost around $5.2 million and cranked out just over 10 teraflops. I remember the total system cost being reported in multiple outlets.

 

The price on those Altix machines is pretty impressive, really. A loaded 512-CPU system sells for a little over $1 million; the NASA system is a cluster of 20 of those.

256 Xserve cluster nodes (for 512 CPUs) would go for $768 thousand, plus high-speed networking hardware, so you're roughly in the same price range as the SGI machine -- but you still don't have the shared memory across 512 CPUs available in the Altix. That difference enables working on a whole range of problems a Mac cluster simply won't be able to solve as well -- they simply don't offer the bandwidth.

 

Note how they are Itanium 2s, Not Xeons. Intels Itanium is a nice processer. Too bad it costs thousands for just the processor.


But.. I wonder what there power and cooling costs are compared to the G5 systems?

 

It isn't too surprising to see SGI using Itanium processors as they were the first vender to widely adopt them early on. What is surprising is that they would venture into the area of supercomputing so soon after selling off Cray.

Well, I still like their workstations. I've never seen or used an SGI Linux/Itanium based system, but I love my three SGI IRIX/MIPS based systems.

 

SGI is focusing more on high performance computing nowadays, rumors of the death of the MIPS/IRIX line have been floating around for years because of the company's financial status during the late 1990s. Why pay for an expensive SGI workstation or server when you can just get a bunch i386s for cheaper and have a cluster of them to do the processing with a System X or Linux front-end?

Jim Clark was right, cheap workstations were the wave of the future, and SGI is paying for the inability to grasp their founder's foresight.

However, IRIX is a very nice Unix OS for being System V (I prefer BSD conventions), excellent scalability, a pretty nice, if simplistic FWMV-based window manager (4Dwm). I bought an old Indigo2 IMPACT on eBay a while back (They're dirt cheap nowadays), I haven't been able to use it much though as I don't have a ethernet port I can plug into the room since my family moved.

SGI hobbyists are rather ignored though, for reasons that are probably purely economic (i.e. Not pissing off corporate customers), as IRIX is insanely expensive and the company made the maintenance release pay-only some years ago.

And if one is wondering where the name IRIX came from, the line of workstations during SGI's heyday was named IRIS, hence, IRIX.

Hmm...anyone want to try to put a Power Mac G5 in an old Indigo case?

 

I'd rather have a Cray@52.4 Tflops.

 

But it's not really the price per processor that's important, it's the cost per teraflop (to simplify it somewhat). And in that equation Apple still comes out ahead. OK so they have 20 of these 512-proc Itanium boxes at $1 million each. That's $20 million and I'm going to add 25% onto that for infrastructure, cooling, communication fabric, etc. I think that's extremely generous, it's probably a lot more than that, but let's go with it (remember you're housing and cooling 5x as many processors, which each probably generate considerably more heat than a G5 chip does). So that's $25 million divided by 42.7 Tflops = $585,000/Tflop for Itanium cluster. Now the Apple cluster cost a total of $5.2 million initially, plus $600,000 for the Xserve upgrade and now scores 12.25 Tflops. 5,800,000 divided by 12.25 Tflops = $473,000/Tflop for G5 cluster.

 

The thing is, that those nubmers, 42.7 TFlops, were based on only 75% of the CPUs running.

If it scales evenly, when they get all the CPUs online they get 53.4 TFlops!!!

Now, what are they going to do with it

D

 

Right... and at that point it's a little cheaper than the Apple equivalent, and still should perform far better on problems involving intermediate results that don't parallelize well.

One of the stated purposes is spacecraft design. Hopefully they'll get the chance to do something better than sink more and more into extending the life of the STS program.

 

I guess IBM's 970 just isn't flop-py enough.

 

Seriously guys, what did you expect? Intel chips are STILL STILL STILL faster than ANY G5!

Sure, G5's are fast (I own one) but Intel is faster. Maybe not for long, but yeah, Intel makes faster chips, both clockspeed and actual performance.

 

As long as my iMac runs Photoshop fast enough to keep me from pulling my hair out... I'm a happy camper!

Let Apple build for us, and let Intel build for the government.
Gee, the government wouldn't know what to do with something that actually worked!

 

I think you need to qualify your claim, these are Itanium 2 chips. Not P4 or Xeons. 2rc, The Itanium has a much lower clock than both the p4 and the g5. But... the Itanium is the Alpha reincarnated. The Itanium is fast, but you must be using intels complier. The Itanium (and the Alpha btw) come from a family of high performance cpus that require the complier to perform alot of "optimizations". Things like instruction scheduling, branch prediction ect. are handed to the complier. Therefore the cpu can spend more space on pipelines, and cache. The xeon, and g5 do not have anything like this. They are "consumer" chips, and most people(developers) can not afford the $10k+ license for a compiler.

btw, the xeons (and p4, the p4 is a MESI, multiprocessor crippled Xeon), are not very good design. there pipelines where designed with one goal; Increase the speed of the processor so we can tell joe consumer that our processor is better.

Also. from what i recall (correct me if i am wrong), The itanium is very hot and power hungry (not to mention the custom ASICs for the internal bus on these monsters) So what are the power, and cooling requirements for these things? are they included in the total cost? I assure you these figures are nontrivial.

 

My G5 is quite stiff actually.

 

Here's the official site

Your link is from a forum, so I have provided NASA's official site which provides additional info.

http://www.nas.nasa.gov/

And to answer your question before you ask, no I didn't look up the link. A friend of mine who worked on the Space Shuttle and the Mars Rover project sent me the link.

As an aside, I did pretty well at math, both in school and college - algebra and calculus were my fortè - but I feel really dumb when I go into his study and page through his maths books on dynamic aerodynamics etc.

 

Sure!

Top-of-the-line P4's, Xeons, and even AMD chips are faster than any G5 to date. I'm not knocking the G5-like I said, I have one. But it is slower than any 3.2+ P4's I've used. (mostly determined by Maya render benchmarks, which is real world enough for me, since I use Maya all the time)

It's no big deal, its just no surprise

 

The first generation Itanium chips were indeed pigs, but the Itanium 2 is in the same ballpark as the G5 (62W is the worst case for I2, typical less).

Really though, these number are pointless. Apple simply do not build massive multiprocessing machines, they're simply not in that business. Apple can offer clusters which can take on some of the same jobs, but they simply aren't directly comparable architectures (that is, different at the system level, not even talking about CPU nitpicks).

 

Revenge of the PowerPC: http://www.research.ibm.com/bluegene/
Once this baby is completed, those Intel-based systems will look very slow indeed...