Opteron + Cell = 1.6 Peta flops
Sorry, Cray, IBM won the bid for the fastest supercomputer with Opteron and Cell. A massive 16,000 Opteron cores ( 2000 x3755 4P servers) and 16,000 Cell CPUs deliver the awesome computing power no FSB based IA32 technology can match. This is AMD64 and Torrenza in action.
Clearly, IBM and AMD have been working with each other for quite a while to interface Cell to the AMD Torrenza de facto industry standard with cache coherent HyperTransport.
AMD sold 8,000 Opteron 28xx CPUs in this deal. At $15,00 CPUs each, that's $12 million, enough to feed 10K workers for four days.
posted by Sharikou, Ph. D. @ 9/07/2006 04:38:00 PM
17 Comments:
Let's start/refresh the rumor IBM to buy AMD. Hehehe...
-Longan-
Interesting to note is that the 1.6 petaflops is either conservative, or DP, which cell is slow at.
1.6 Petaflops = 1.6E15 Flops
1 cell in SP ~ 250 GFlops = 2.5E11 Flops
For 1.6 Pflops peak: 6,400 cell processors. This is alot less than 16,000. Either that means the predicted performance is 1.6 PFlops, or cell is going to be alot slower than original estimated.
This is great news and shows who's really got the power here.
Intel is in 60.4% of the Top 500 Super computers as compared to 16% for AMD. I guess they can let one go -:)
Top 500 Super Computers
Both sides of the story always...at:
sharikou180.blogspot.com/
I am wondering if someone could explain this for me...
2000 4P servers = 8000 Opteron processors, ok
Here is the main part of the question, isn't Torrenza a "socket" orientated co-processor?
Where are the 16,000 Cell processors going?
Thanks in advance :)
Where are the 16,000 Cell processors going?
I would like to know the same thing.
If I am not mistaken, the finalists were Cray, IBM and Sun.
So there was never in doubt Opterons would be used.
It will be Opteron plus proprietary processors.
Has Dell join the Hypertransport consortium yet? Soon? :)
enumae:
According to the press release, IBM is using 16k Opteron cores. Let's look at each server - 4P hence 4*2=8 Opteron cores in each, or 16k/8=2000 x3755 servers (http://www-03.ibm.com/systems/x/rack/x3755/specs.html).
The Cell processors are in separate servers - IBM BladeCenter H systems. I couldn't find them on IBM's website. Are they on sale yet? I guess not.
I don't think IBM is using the Cell processors as a co-processor on the x3755 motherboard itself. However, the x3755s and the H series BladeCenters are probably connected by some sort of HTX slot/connector since they are claiming in the press release that "This is an excellent demonstration of Torrenza in action".
We know that Opterons and the Cell processors are on separate motherboards and Torrenza is being used. Hence they must be using the HTX slot/connector as this would be the only other way (if my understanding is correct) to use the Torrenza technology in this context other than connecting all the x3755 Opteron servers together.
So if they are connecting the x3755 servers using Torrenza, why not the interconnects between the x3755s and the H series BladeCenters as well?
http://www.theinquirer.net/default.aspx?article=34231
BUAJAJAJA
intel again cheating and using guerilla marketting to make AMD look bad.. intel is seriusly looking very pathetic.
kicks of a almost-drown company anyone?
4P server has 4 CPU socket and 1 Hypertransport Slot (HTX).
Since each Cell processors contains 8 CPU cores, so there is 2000 x 8 or 16000 Cell CPU cores.
Official HTX Information
Well, they'll actually have to build it to see how well it really works, but it's promising.
"180 Sharikou said...
Intel is in 60.4% of the Top 500 Super computers as compared to 16% for AMD. I guess they can let one go -:)"
The Top 500 have some legacy systems..
Look at the top 10.
http://www.top500.org/lists/2006/06
From what I can tell.. it is close: 2 Intel vs 2 AMDs
Expect the Cray XT3s to climb up the list when they get their dual core upgrades :D
And with this new Cell/Opteron deal?
Opteron needs Cell and Cell needs Opteron to hide their defeciencies. That's why they had to use both the CPUs.
IBM has thrown money on CPUs to get to No.1 spot. There is nothing in it that show superiority of either Opteron or Cell.
You people are so far off base. This is a collection of off-the-shelf hardware-- IBM "System x" model 3755 Opteron servers plus IBM Bladecenter H servers loaded with Core B.E. blades. The interconnect is not specified but is probably Infiniband.
This deal has NOTHING to do with HyperTransport links between the Opterons and the Cell processors, and NOTHING to do with Torrenza.
C'mon, read the press release, why don't you.
http://www-03.ibm.com/press/us/en/pressrelease/20210.wss
This is the Bladecenter H press release:
http://www-03.ibm.com/press/us/en/pressrelease/19198.wss
Let's do some math. According to the Inquirer article here (http://theinquirer.net/default.aspx?article=31836), 2 core 2.8GHz Opteron has an Rmax of 11.8Gflops. Giving the benefit of the doubt with 100% scaling efficiency (not likely, the article predicted 77%/double socket), that means 16,000 2-core Opterons would yield... 188.8 Gflops. That would mean the 16,000 Cells are then accounting for the remaining 1.4 petaflops of performance? Even if you can extract double the performance from the Opterons, they still only account for less than 25% of the performance. The realistic case with scaling inefficiencies is ~145Gflops, or 9% of the system performance. Either I am missing something, or I don't see how this is a big deal for AMD, other than they get a 1 time deal for 16k CPUs, which is probably less than a 1/2 day's sales to HP. Seriously, what am I missing here?
Nova kid is an amateur. His numbers are useless. But, your multiplication skill is off. Suppose one 2.8GHZ Opteron has 10G flops. 8000 of them yield 8x 10^4 Glops. or 8 x 10^13 flops, or 80 Tera flops.
A petaflop is 10^15 flops. or 1000 teraflops. Opteron is not contributing to FP performance here at all. It's Opterons massive I/O scalability that is important.
"If I am not mistaken, the finalists were Cray, IBM and Sun.
So there was never in doubt Opterons would be used."
What made you no doubt that IBM would use Opteron instead of its own (non-Cell) Power processors? ;-)
Post a Comment
<< Home