“VT Xserve G5s to be at least 2.4GHz: The usual suspects say there is a very good reason for the timing of Virginia Tech’s switchover to Xserve G5s in the April/May timeframe – that is when Apple has said they will have faster dual PowerPC 970FX-based Xserves with CPUs running at no less than 2.4GHz. At that time Apple expects to be shipping PowerMac G5s with Dual 2.6GHz 970FX processors, and the only reason the Xserve lags slightly behind is the still considerable heat production of the 970FX. Not long after, Apple will be in the final stretch of development and manufacturing of the Powerbook G5 which could run as fast as 2GHz,” MacOSRumors.com reports.
“Due to the additional 800MHz of additional combined clock speed per node along with the other performance benefits of a cluster that will be 3X denser than the PowerMac-based supercomputer, the new system ought to be considerably more powerful even with the same number of nodes as before. The real killer about this latest wave of reports is this, however: rather than the 1,100 nodes in the PowerMac cluster, the extra space freed up in the Xserve system might allow VT to scale to 1,600+ nodes without too drastic a financial outlay on VT’s part due to a very kind package deal offered by Apple…. Most impressive. We will just have to wait and see how close to the mark these reports prove to be,” MacOSRumors.com reports.
Full article here.
Related MacDailyNews article:
Forrester Analyst: Regardless of Virginia Tech ‘Big Mac’ supercomputer, IT pros will ignore Apple products – January 27, 2004
They are going after Earth Simulator =) Great!
The Most Powerful Super Computer In The World
” width=”19″ height=”19″ alt=”smile” style=”border:0;” />
Gigantic X (New name)
Nice. Oh Joy!
That is marketing money well spend =)
If only this hadn’t come from MOSR.
Oh come on. Not MacOSRumors again!!! Let’s not even go there.
Gosh, Macosrumors.com is SO reliable. What do they have, like a 5% accuracy rate?
This one seems like sort of a no-brainer, so even Macosrumors could make this one up and feel like they have about a 50/50 shot at being right. The problem is that the switchover won’t happen until April/May for reasons other than when Apple is going to ship them. The switchover will happen then because of the work that the cluster is currently doing, the time it will take to build everything out and to move out the existing PowerMacs, and a plethora of other details. If the Xserves didn’t ship until April, they wouldn’t switch over until June.
Now, as for the other parts of the rumor…
Since Jobs said the G5 would be running at 3GHz by Summer, then they better hit 2.6 by May. Duh.
Since the 970FX consumes less power and produces less heat (and is smaller) than the same clock rate for the G4s in the current batch of PowerBooks, then a migration to the G5 for the PowerBook very soon is logical, too. The first PowerMac G4(350MHz) was announced 16 months before the first PowerBook G4 (400MHz). The development of the 970 processors has been happening much more rapidly than the early G4s, so it seems unreasonable to think that the first PowerBook G5 would even be 12 months behind the first PowerMac G5. We’ll see.
Wow…now that would be freakin amazing!
This actually makes sense. If Apple buys that code from Virginia Tech and Varadarajan with certain amoun of xServe G5’s it would be nice deal for both. MacOSRumors is not important here. Virginia Tech has something that Apple want’s and Apple has something that Virginia Tech want’s. This is excellent marriage! Whoo!
Hey Finland guy, there is no way they are going after the Earth Simulator. A quick look at the numbers would make that obvious. The Earth Simulator is over 3X powerful than the VT cluster.
Don’t post BS.
MacOSRumours makes a blanket statements that simply is not true:
“At that time Apple expects to be shipping PowerMac G5s with Dual 2.6GHz 970FX processors, and the only reason the Xserve lags slightly behind is the still considerable heat production of the 970FX. “
The 970FX does not have a heat issue even at 2.6GHz, these are 1.7″ high units and very deep, compare this against 1.5″ notebooks running P4 3GHz processors that emit 90W!
The Xserve G5 2GHz 970FX is faster than the 2GHz 970. It uses 30% faster transistors, which are also closer together by nature of the 90NM process, so the latency is a lot lower. So right now, the Xserve doesn’t lag the PowerMac, it is faster. A more acceptable reason why the Xserver is at 2GHz is that these are the first run 970FX, so IBM is fine-tuning their new process and 2GHz produces the best yields and the lowest cost for Apple. It’s a smart strategy that puts the 90NM process out there right away as IBM ramps up production and yield for the 2.4 and 2.6 GHz+ processors.
MV
They can do it.
If xServe takes only 1/3 of the space=You can have 3 times more computers in the same space and it will still be a cheap super computer. Apple could easily buy Virginia Tech’s special super cluster programs with $10 000 000� ie extra 2200 xServe G5’s. Or Virginia Tech has now investors who has the money.
Let’s look at the numbers shall we:
3*10,28TFlops is=30,84Tflops. That falls 5,02TFlops short from Earth Simulator.
If xServe gets more speed up to 2,6GHz=30% more.
30,84TFlops+30%=40,092TFlops. That is 4,232 TFlops MORE than Earth simulator.
If xServe gets more speed up to 3Ghz=50% more.
30,84TFlops+50%=46,26TFlops. That is 10,4TFlops more than Earth simulator.
Better look at the numbers makes it more obvious so don’t post BS
Ahhhh…
But if they shrink the current cluster down to 1/3 the size, can’t they make a run at the #1 supercomputer by filling up the place with xServes? Plus the ones they are going to use are going to be fast then the current PowerMac G5s, right?
As stingerman points out Virginia Tech’s xServe G5 Super Computer could be extra 30% (because of the 970fx) faster. Even though I did not count that in System X is would be faster.
Finland…
Your logic stands if the performance increase is linear. So, 9.34*N tflops (assuming 10,280/1100 nodes=9.34 each). In your case, 9.34*N*3= 30,840. I HIGHLY doubt this will be correct because as you add more nodes communication and other such cperformance hits rise.
You are also assuming that a 30% speed boost in the CPU power will translate to a realtime 30% performance boost. This just won’t happen.
For Apple to make the world fastest super computer would be extremely good PR. Steve knows that. Virginia Tech knows that. You know that. I know that. MDN knows that. Everybody knows that.
Right?!
jeff
They can waste 9TFlops and it still would be the fastest super computer in the world. Though I don’t believe that would happen.
The point is that it is bossible and worth trying.
February marks the real 6 Month point since the dual 2GHz G5 started shipping. Since Apple is modifying the Case to accommodate another optical drive and additional SATA drives due to the reduce cooling requirements of the 90NM chips (processors and ASIC controller), Apple is waiting till February to roll-out updates.
The ES is based on 5,120 (640 8-way nodes) 500 MHz NEC CPUs.
I believe that 6600 (3*2200) 2,6GHz G5 processors would beat that. Actually they don’t even need so many.
Jeff, what has been incredible is that the increase in GHz has been linear. In this case even slightly better. Why? Because the the FSB grows as well. In the old days we would measure a processor by both the internal and external clocks. Then when external clocks (FSB) froze, we only referred to the internal clock. With the G5 architecture Apple has brought back a more accurate performance measure. With every 2MHz internal increase, there is a 1MHz external increase. So we get a more linear increase, as the Internal overhead is accommodated with twice the clock of the FSB.
In addition, the switch to not only a 90NM processor but more efficient components. IBM went from their CU-11 components to their newer CU-08 components, which overall are 30% faster. So we should expect a very linear performance increase.
What is next, well, it appears Apple may jump to a 1:1 ratio for the processor to processor communications. This will result in seamless cache pooling between the processor. This will probably happen when IBM delivers to Apple a dual core 970 (maybe 980 by then). Since Apple mainly sells dual processor PowerMacs, it would actually reduce their costs to spec out a new processor with dual cores and maybe just one Altivec2 (the rumored 256-bit Altivec Unit). But, that is just too exciting to dwell on.
Thank you stingerman. That was score!
(My english get’s worse everytime when I get excited. So sorry for the typos)
We are truly living intresting times. Mac is really sexy platform at the moment and the future.
most powerful bullshit engine: 100 terashovels
MacOSRumors.com
One guy from Finland wrote:
“Better look at the numbers makes it more obvious so don’t post BS”
I’m afraid you’re the one posting BS. Inexpensive clusters built with commodity hardware — be they Intel/AMD/PowerPC-based — just cannot compete in terms of sheer processing power with a real supercomputer like the Earth Simulator.
One could compare the E.S. to a massive 40-ton trailer that can reach 200Km/h. The fact that you can build a SUV that can also reach 200Km/h doesn’t mean that the SUV is as powerful as the trailer…
Compare the memory bandwidths: with two banks of DDR400 memory, on a dual G5 machine, the average memory bandwidth to each CPU is 3.2Gbytes/sec — i.e. just enough to send 400 million 64-bit floating point numbers per second to the CPU. With memory-intensive applications requiring huge datasets, the slow memory bandwith is going to really throttle the CPUs.
With the E.S. on the other hand, each of the 5120 CPUs has a memory bandwith of 32GB/s — i.e. *an order of magnitude* faster than the G5. On large computation problems that don’t fit in a cache, the memory bandwith is really going to make a difference.
With problems requiring large shared datasets involving data transfer across nodes, commodity hardware-based clusters using Infiniband technology will have a bandwidth limited to 10Gbits/sec (before taking overhead like 4b/5b encoding and error detection/correction into account,) and the network latency will also be pretty significant as the data packets might have to travel through several Infiniband switches before reaching their destination.
With the E.S., on the other hand, each of the 640 8-CPU nodes has a *direct* one-hop link to *any* of the other 8-CPU nodes in the cluster. The nework latency is thus kept *very* low.
Each of the E.S.’s 8-CPU node connects to this fully-meshed fabric with a 128Gbits/sec link. It’s interesting to note that even these “long-distance” inter-node links are already several times faster than the CPU-RAM bandwidth achievable with 64-bit DDR400 memory as used e.g. in the G5 Macs…
Can MDN not refer to MOSR please? Or at least if MDN does, put a disclaimer that it must be taken with a couple tons of salt.
Hmm..
So how much faster is 3300 G5’s is compared to 1100 G5’s?
You say that 3300 G5 cluster is slower than 1100 G5 cluster?
I don’t think so.
Actually. There is a task for MDN. Go interview Varadarajan and ask what he thinks. Could G5 cluster beat ES?