News: Itanium Montecito breaks performance record

[jjcom]

Intel Corp. said this week that its forthcoming Itanium processor code-named Montecito with two processing engines delivers world’s record performance in LINPACK benchmark, which measures floating point speed. The claim should prove the power of dual-core Intel Itanium chip and create rush around Intel’s dual-core products.
http://www.xbitlabs.com/news/cpu/display/20050708133307.html

[[XC] moddolicous]

This looks promising. If more companies upgrade their comps, then that means many itaniums will be up on ebay soon. Maybe some 1.6lv D1 xeons will be for sale soon also on ebay.

[Hans.Gruber]

system with four dual-core Itanium processors exceeded 45 GFLOPs (gigaflops)
That's slower than two dual core Opterons... 2 x Opteron 275 @ Linpack (http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_8800~96867,00.html)

Can anyone explain this ?

[MetalStorm]

That's slower than two dual core Opterons... 2 x Opteron 275 @ Linpack (http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_8800~96867,00.html)

Can anyone explain this ?

Itaniums suck perhaps?

[StrikeRTM]

Suck's perhaps?

Well who said anything abotu the exact model. Mayeb that was an entery level model and the ywere happy about the result. It is possible there will be models, it is not released ether, so in the end it mgiht have much more in it.

But everything is in a level of spectacualtions, so there's nothing to be shure about there.

[sluflyer06]

well on the AMD side a node would be a single dual processor system...so 2 nodes= 4 dual cores = 29.65GFLOPS/35.2GFLOPS Peak.

so 10GFLOPS behind the Itanium.

[agenda2005]

well on the AMD side a node would be a single dual processor system...so 2 nodes= 4 dual cores = 29.65GFLOPS/35.2GFLOPS Peak.

so 10GFLOPS behind the Itanium.

You surely have no clue what you are talking about. Go back and check the AMD website for the benchmark and below it click on the configuration information. They use opteron 275 systems with 2 cores and 2 nodes/core (four processos in total) Here's what X-bit labs say;

Using the LINPACK benchmark, a system with four dual-core Itanium processors exceeded 45 GFLOPs (gigaflops), a measure of computer speed where a gigaflop is 1 billion floating-point operations per second.

Take note of the word "four dual-core Itanium processors"

Here is the simple comparison

four dual-core Itanium processors (8 processors) = 45.xx GFLOPS
two dual-core opterons (4 processors) = 58.34 GFLOPS

Even though we still assume they are four on four, opteron still wins!
You can now take your pick.

[Cooper]

Heh you missed one part guys:
Intel Claims Dual-Core Itanium to Deliver World’s Record Performance
:D so that`s the answer. And who would spend $xxx on Montecito, when you have nice, cheap and well performing Opteron 2xx . Does Intel have to say anything about this ? guess not :)
I also like this part:
We are approaching the ability to reach a TeraFlop in as few as a 20-server system cluster and helping to dramatically increase the affordability to the scientific community.” :lol:

[sluflyer06]

opterons are undoubtedly awesome processors and if I ever need a workhorse server it'll be AMD all the way...however Scientific Applications CRAVE onboard cache...so the Itaniums do pretty damn good.......also something alot of people don't know is that you can get Xeon's with 4MB+ of cache they are just very hard to come by.

[freecableguy]

You can get Xeons with 8MB of onboard cache actually....

[sluflyer06]

yea thats what the "+" was for because I couldn't remember exactly how high it went...thanx cableguy

[camel]

You surely have no clue what you are talking about.


Actually, he was right. ;)

This bit of your post is wrong:

two dual-core opterons (4 processors) = 58.34 GFLOPS


Look carefully at the AMD chart and you will see "2P systems" (in the top right). The figure of 58.34GFLOPs is for a mini cluster of eight dual core opterons, for a total of 16 individual opteron cores. There are eight actual Opteron 275 chips, like this:

4 Nodes * 2 Processors per node * 2 Cores per Processor = 16 Cores

If you look at "Configuration Information", you will see that the 4 separate servers are connected with Myrinet, a kind of high performance network card (see http://www.myri.com )

Basically, when AMD (or anyone) says node, they are referring to a complete server box, not a processor (chip you can buy) or a core. In this case, a box with two CPUs - either two Opteron 275s (the dark green bars) or two Opteron 248s (the light green ones).

[sluflyer06]

Thank You Camel....the graphs and explenation are quite easy to read and understand so AGENDA2005 my friend I think you should look at the data more closely before you make yourself look like an @$$ again and tell me I don't know what I'm talking about when your dead wrong.

your pretty close though...add 12 more cores to your post and you'll get to the 58gflops/s

[Hans.Gruber]

Actually, he was right. ;)

This bit of your post is wrong:


Look carefully at the AMD chart and you will see "2P systems" (in the top right). The figure of 58.34GFLOPs is for a mini cluster of eight dual core opterons, for a total of 16 individual opteron cores. There are eight actual Opteron 275 chips, like this:

4 Nodes * 2 Processors per node * 2 Cores per Processor = 16 Cores

If you look at "Configuration Information", you will see that the 4 separate servers are connected with Myrinet, a kind of high performance network card (see http://www.myri.com )

Basically, when AMD (or anyone) says node, they are referring to a complete server box, not a processor (chip you can buy) or a core. In this case, a box with two CPUs - either two Opteron 275s (the dark green bars) or two Opteron 248s (the light green ones).


Configuration says:

"AMD Opteron™ processor Model 275 and 248-based system: Hardware: AMD “Quartet” internal development platform (not publicly available). Processor: Qty. (2), L1 Cache 2x 128KB (64KB data + 64KB inst), L2 Cache 2 x 1024KB, Memory: 8GB total. Hard Disk: 15 x 36GB U320 SCSI. Operating System: SuSE Professional 9.2 2.6.11-rc3 kernel patch-9962 Release 20030321. Libararies: Goto, mpich-gm mpi. Network: Myrinet dual-port card."

It doesn't say anything about it being 1 node, but it also isn't very clear at all..

So can you give us a link to a source of your information ?

[sluflyer06]

that is the setup for a node....

[camel]

It is a little unclear, but when people talk about clusters, the language normally is:

node: a complete computer system, running its own copy of the OS (Linux or whatever)
?P or ?-way system: a computer system with ? many CPU sockets

Note how in "Configuration Information" it talks about system but on the performance graph in the top right it talks about systemS. The numbers on the left (1 node, 2 node, 4 node) are counting how many systems are being used, connected together with myrinet.

The real giveaway is this bit, though:

Configuration says:
mpich-gm mpi. Network: Myrinet dual-port card."


Myrinet is used for hooking up computers in a very fast way (much, much quicker than ethernet), and it's commonly used in clusters. But if you still had your doubts, "mpich-gm mpi" should tell you what's going on. MPI stands for "message passing interface", it's one of the most popular ways to write programs for highly parallel supercomputers, like clusters. Check out the LAM MPI website for more info: http://www.lam-mpi.org/

(AMD used another kind of MPI called MPICH, but the idea is the same)

[Xassius]

agenda2005 has been known for looking like an ass. :D