Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Virginia Tech Supercomputer Ranks #4 [updated to #3]
- Thread starter MacRumors
- Start date
- Sort by reaction score
Cheap-tech
The proof that Macs aren´t expensive. And what´s the maintenance cost? Close to nothing compared to the other top-ten I would assume.
The proof that Macs aren´t expensive. And what´s the maintenance cost? Close to nothing compared to the other top-ten I would assume.
If I remember rightly the US Navy had ordered a number of Macs [Possibly Xserves, I can't remember?!] to which it was quoted they wouldn't be installing X and indeed would be whacking one of the many flavours of Linux on there...
PS. This is my FIRST post... whoooh!!
IMO this is a huge achievement for both Virginia Tech & Apple. To come in at #4 with these preliminary results is amazing for a supercomputer costing just US$5m!!! It is great PR for Apple and the G5 PowerMac that the CPU, and in fact the entire machine that is available for everyone to buy and have on their desktop at home, can hold its own in the supercomputing world.
I think it is great that a Mac cluster is likely to be in the top 5. I would be even happier if many groups started making smaller clusters of Macs and knocked a bunch of those HP clusters off the list. It's great to be on the list, but claiming just one out of five hundred spots isn't so great. I guess it's nice that quite a few are made by IBM.
I really like the sound of "The Earth Simulator"
don't know why just sounds impressive.
The G5 cluster sounds like a good way to go in many ways. I think Apple should concern themselves with producing medium power Supercomputers to allow research projects to utilise them. A new generation of more affordable but still powerful SCs could push science IN GENERAL forward a little bit faster. But there will be areas, such as weather prediction, that will require the sheer power of SCs like the Earth Simulator.
don't know why just sounds impressive.
The G5 cluster sounds like a good way to go in many ways. I think Apple should concern themselves with producing medium power Supercomputers to allow research projects to utilise them. A new generation of more affordable but still powerful SCs could push science IN GENERAL forward a little bit faster. But there will be areas, such as weather prediction, that will require the sheer power of SCs like the Earth Simulator.
I read somewhere (mebbe on MacRumors even) that this wasn't the full 2200 processors but 2100 or so. Still, a hand-full extra won't take it to number 3. Maybe after a few more optimisations?
Still, great to be top 5, top 3 would've been the icing on the cake
biscuit
Still, great to be top 5, top 3 would've been the icing on the cake
biscuit
Well, seems that people are still comparing the Virginia cluster to the old list (June). Let's see how well the other supercomputers that have been added in the meantime perform... I thought IBM also wanted to launch a supercomputer that they think will make top 5?
It's the network !!
Scaling on a massive cluster is largely a function of the network interconnect, it needs to be very fast and very low latency.
The VT cluster uses InfiniBand, an Intel-developed high speed interconnect fabric.
The Xeon (and Alpha) clusters in the top 500 list are using slower interconnects, which results in lesser efficiency and poorer scaling.
Y'all should prepare yourselves for Xeon and Opteron clusters using the InfiniBand networking.... You'll see that the network is what's special about VTech's cluster, not the CPUs.
Scaling on a massive cluster is largely a function of the network interconnect, it needs to be very fast and very low latency.
The VT cluster uses InfiniBand, an Intel-developed high speed interconnect fabric.
The Xeon (and Alpha) clusters in the top 500 list are using slower interconnects, which results in lesser efficiency and poorer scaling.
Y'all should prepare yourselves for Xeon and Opteron clusters using the InfiniBand networking.... You'll see that the network is what's special about VTech's cluster, not the CPUs.
Tell your friend not to be stupid, get the iBook. I have a G3 900MHz and it's every bit as fast as my best friends 867MHz 12" Powerbook (and it cost about $500 less at the time). Now that it has a G4 why bother with the 12" Powerbook? The only real difference is the Powerbook keyboard is nicer.
This is a prelude to G5 Powerbooks by the end of the year in my opinion (maybe after the first of the year).
Got it. Although not immune from error (duh.) the great dailies (The Wall Street Journal, The New York Times, The Washington Post, The LA Times) usually do not print rumors. Although there is the "unnamed sources" stuff. Moreover there is an editing process that goes on in any newpaper which results in not presenting the unmediated/unedited "facts". Although, recently, at least in the political arena. the fact that there are "rumors" has actually become news so that the press begins reporting on the fact that there is rumor and that becomes a news story. I recall, the later disproved, "ossuary of the brother of Jesus" and the "Bill Clinton murderd Vince Foster" stories. I'm sure that there are other rumors and facts that don't get published immediately by the respectable press. Of course, I read online just this week that review from the WinDoze people on iTunes was mixed: I guess that was true--only 87% loved it.
If you'll notice...
If you look at the Rpeak vs the Rmax of all the other supercomputers in the Top 5, you'll notice that they are WAY closer to their Rpeak than the score for the VT Cluster is. For comparison the current Top 5 are at 87.5%, 67.7%, 69%, 59.4%, and 73.1% efficiency. In fact, in the Top 40, none of them go below 45% efficiency. The VT Cluster, given these numbers, is at 42% efficiency.
I'm willing to bet that this number comes from one of the first times they've run the benchmark on the cluster as a whole. They haven't shaken all the bugs out yet, and the certainly haven't optimized the benchmark for the cluster as a whole if this is true
Assuming the VT Cluster follows the trend of it's fellow super computers in the Top 40, it should perform at at least 7.92 Tflops once all the bugs are shaken out - and that is still less than 50% efficiency. If they can get efficiency to the level of the Top 5 computers, it should post over 10 Tflops. Personally, I think that they can easily move up to 3rd and score over 45% efficiency - 3% is well within the limits of optimizations.
The real question to ask is what other Super Computers have been built and are in the running for Top 5 spots this year...
If you look at the Rpeak vs the Rmax of all the other supercomputers in the Top 5, you'll notice that they are WAY closer to their Rpeak than the score for the VT Cluster is. For comparison the current Top 5 are at 87.5%, 67.7%, 69%, 59.4%, and 73.1% efficiency. In fact, in the Top 40, none of them go below 45% efficiency. The VT Cluster, given these numbers, is at 42% efficiency.
I'm willing to bet that this number comes from one of the first times they've run the benchmark on the cluster as a whole. They haven't shaken all the bugs out yet, and the certainly haven't optimized the benchmark for the cluster as a whole if this is true
Assuming the VT Cluster follows the trend of it's fellow super computers in the Top 40, it should perform at at least 7.92 Tflops once all the bugs are shaken out - and that is still less than 50% efficiency. If they can get efficiency to the level of the Top 5 computers, it should post over 10 Tflops. Personally, I think that they can easily move up to 3rd and score over 45% efficiency - 3% is well within the limits of optimizations.
The real question to ask is what other Super Computers have been built and are in the running for Top 5 spots this year...
Well, the VT cluster posted some very prelim results using 128 nodes and was projected at only 7.055 Tflops when complete. With all 1100 nodes it actually scored (prelim) 7.410 Tflops. So it looks like extra nodes DID help a lot. But as AidenShaw pointed out, it is often more a function of the communications fabric than the number/speed of the CPUs themselves (this of course depends on what the cluster is doing - if it's a massively parallel job then the com fabric may end up idle for most of it).
If they did double the size of the cluster, it probably could challenge the #2 slot and win it.
Edit: Oops, wrong number of nodes in the final cluster.
Re: iBooks updated
Please stay on topic
Why are either of you brininging this up in a thread about the Virginia Tech Supercomputer!?!
Please stay on topic
Re: Cheap-tech
Sushi
What is really cool about this system, is after a year or two when the G5 chips are much faster, they can upgrade the entire system easily by just purchasing new G5 Macs. At the same time they can sell off the old ones offsetting the cost of the upgrade somewhat. Sweet.
Sushi
So the final #'s are due in November, should we expect them to be much higher?
Sure there are lots of processors coming that will be even faster but for now if it turns out that the Mac is #4, I think that's quite good, being that it is the first Mac on the list?
6 months from now should be even more interesting
daniel
Sure there are lots of processors coming that will be even faster but for now if it turns out that the Mac is #4, I think that's quite good, being that it is the first Mac on the list?
6 months from now should be even more interesting
daniel
actually I think the original 128-node sample put it at 80% of the theoretical maximum of 17 Tflops. That's where the 14 Tflop figure came from (17 x 0.8 ~= 14). So this latest result of 7.4 is drammatically lower than that earlier projection. As another have noted 42% efficiency is A LOT worse than the other top 5. So while the cost-performance ratio is excellent, it's still a little disappointing.
The 7.4 teraflops number you are tossing around is indeed the maximum but it is not the peak. The peak is around 17 teraflops so everything is scaling exactly as they expected.
well -- they did have a little incident where they had a writer making up stories and an editor who was changing statistics...