Itanium Vs Xeon

[dobby]

When woudl you use an Itanium Processor over an Xeon and who is it aimed at (obviously not the Home nor small server space.)

thanks alot.

 

[hsarc]

Itanium is designed for a large server that requires a true 64-bit processor for large memory address and lot's of processor power. In fact because of it's design it does not run 32-bit Apps very well at all. Usually companies will buy this processor on a server for a specific application such as a large database, analytics, anlysis, etc.

 

[Heyyou27]

The Itanium line is designed for the enterprise server market; it runs a completely different instruction set than Xeon or your standard desktop CPUs and requires its own varient of Windows server or Linux.
http://www.microsoft.com/servers/64bit/itanium/overview.mspx

Itaniums are also very expensive in comparison; Itaniums generally run from $1000 to $4000 for a single CPU.

 

[Amiga500]

When woudl you use an Itanium Processor over an Xeon and who is it aimed at (obviously not the Home nor small server space.)

Xeons don't work well even in 2 socket systems for high memory bandwidth applications.

Itanium still has problems with the FSB and mem bandwidth with relatively poor scaling for high memory apps.



I have only ever used Itaniums in HPC Clusters - I don't know how good they are in the server space, but I'm not overly impressed with them for HPC work.

 

[jimmysmitty]

Xeons don't work well even in 2 socket systems for high memory bandwidth applications.

Itanium still has problems with the FSB and mem bandwidth with relatively poor scaling for high memory apps.



I have only ever used Itaniums in HPC Clusters - I don't know how good they are in the server space, but I'm not overly impressed with them for HPC work.

That is until the new Itanium with the IMC and QPI comes out. But it will still be purely IA64.

 

[leon2006]

Apples and Oranges....

 

[Amiga500]

That is until the new Itanium with the IMC and QPI comes out. But it will still be purely IA64.

True, they should help with scaling.



But, truth be told, I don't see IPC benefits of the architecture big enough to offset its cost or its clock speed limitations in the work I'm doing. A dual core run on a 1.4 GHz Madision is significantly slower than a dual core run on a dual socket Opteron 250.


*Since Itaniums are always clocked much, much lower than equivalent priced x86 chips, it is a valid comparison of sorts.

Tukwila only goes to 2 GHz.... and rumours abound that all cores cannot run flat out at the same time.

 

[physiker]

I've used a large cluster of Itanium-based machines for computational physics. They have an architectural advantage if you need the precision of 64-bit arithmetic. They are very close to their peak Gigaflop rating for regular algorithms like matrix factoring; if that describes your problem they are cost-effective -- they drop off by maybe a factor of two for general calculations. The individual processors may be expensive, but (a few years ago) it was hard to beat the price/performance of a full system, or just get enough aggregate computational power in other architectures without buying many, many, processors.

I have a desktop dual-Xeon, and it's fast for day-to-day apps (e.g. Matlab), and I'm impressed by what it can do, but it's several times slower than an Itanium for running computationally intensive problems.

 

[dobby]

Ok, thanks alot for all this input. it jsut sort of clear up what the hell its for. thanks again

 

[Amiga500]

They have an architectural advantage if you need the precision of 64-bit arithmetic.

Ahh, yes yes.


That is the advantage of the Itanium in a nutshell.


I can get away with 32bit precision, so while I do need the memory registers, I don't need the numerics. So I don't really take full advantage of the architecture.

 

[jimmysmitty]

Ahh, yes yes.


That is the advantage of the Itanium in a nutshell.


I can get away with 32bit precision, so while I do need the memory registers, I don't need the numerics. So I don't really take full advantage of the architecture.

So does this mean that all the x86-64 does is give the ability to support more memory but only Itanium, Intels IA-64, is true full blown out 64bit supporting every last benefit that comes with it?

If so that would just suck. I would have prefered to have gone to a new type of CPU that has all of the features. I am sure one of them will look into a 128bit CPU, although we wont need it anytime soon. We can call it x86-64-128.

 

[Amiga500]

So does this mean that all the x86-64 does is give the ability to support more memory but only Itanium, Intels IA-64, is true full blown out 64bit supporting every last benefit that comes with it?

No.

x86-64 is an extended 64bit instruction set on top of the 32bit stuff.

Itanium is 64bit from the ground up.


Its kind of like comparing a Honda Civic someone has tuned the engine a bit in, and a NSX. The Civic will be as quick (if not quicker) in some respects, but not come close in others.

 

[yomamafor1]

When woudl you use an Itanium Processor over an Xeon and who is it aimed at (obviously not the Home nor small server space.)

thanks alot.

Itanium are mostly used in enterprise servers, and scientific communities that require large MP solution.

 

[leon2006]

Itanium hardware is a lot better because its not bag down by the architectural requirement of X86.
Apples and Oranges...

Itanium is more of RISC(Reduce Instruction Set) type CPU and X86 is a CISC(Complex Instruction Set Computing)

Itanium can emulate x86 through software but it will be slower. Thats the reason Intel can't make Itanium x86 compatible. Intel will have to add more tansistors/circuitry to make Intanium x86 compatible.

RISC Type Processors:

Power PC (IBM , Freescale, PS3-Cel CPU designed by IBM)
SUN SPARC Processors

CISC TYPE Processors:
Xeon x86
Pentium Series X86
AMD X86
MCS 51/96 Micro-controllers

RISC Advantage
1) have fewer instruction sets
2) Run at a higher clock rate
3) Lower Transistor Count
4) Higher efficiency.... Multiple Instructions Execution per Clock Cycle.
5) Ist not bug down by the DEEP-INTRUCTION PIPELINE needed by CISC such as X86 Architecture.


Big Iron Machines for Scientific and design applications mostly use RISC. The LOS_ALAMOS labs that simulate nuclear explosion. Circuit Simulation & Synthesis that takes months to compute normally use RISC. These supercomputers have > 10K CPUs working together. Common software is UNIX/Linux Operating system, Design , Mathematical and Engineering softwares

CISC is mostly for office application, business, banking, games. CISC servers are cheap than RISC thats why its very popular and widely use. And of course MS-Windows, and MS-Office is probably 85% of the main application.

 

[MU_Engineer]

When woudl you use an Itanium Processor over an Xeon and who is it aimed at (obviously not the Home nor small server space.)

thanks alot.

Itaniums are largely aimed at the high-performance computing market that uses hundreds of CPUs in parallel. Some are also being used in big render farms and for huge databases. Itaniums are known for their powerful FPU and have an extremely wide (11-issue) core. They will do very well at number crunching on heavily-tuned algorithms in massively parallel situations.

Xeons are simply general-purpose CPUs that are "allowed" to do multi-socket bus setups. They may or may not have extra L3 cache. They are much faster than Itaniums on most applications as the Itanium is a strictly in-order CPU while the Xeon can do out-of-order instruction execution. Code running on Itaniums absolutely require good tuning to run well, but you can use a mediocre compiler to compile mediocre code and not spend much time on optimizations and it will run decently on Xeons. Xeons are also x86_64, which lets them run Windows programs. Yes, I know, there is an IA64 Windows version but there are about ten programs that run on it and ten people that use it, so it's insignificant.

 

[Shadow703793]

Why don't I throw in CUDA and a 9800GX2 for the Xeon/Itanium mix ? . The GPUs can actually calculate Floating point and smiler operations faster than a CPU and cost less.

 

[MU_Engineer]

Why don't I throw in CUDA and a 9800GX2 for the Xeon/Itanium mix ? . The GPUs can actually calculate Floating point and smiler operations faster than a CPU and cost less.

GPUs are most certainly fast, but they are specialized pieces of equipment and have certain limitations that CPUs don't have. A really big one for HPC applications is memory. A mid- to upper-level GPU usually has 256 MB or 512 MB of onboard (framebuffer) memory. That memory is very fast, but any data the GPU core would need that is not in the framebuffer has to be fetched from the main RAM by the memory controller, passed from the memory controller to the northbridge or southbridge (dependent on chipset), and then sent over the PCIe bus to the card. That is much slower than just fetching the data from the main RAM directly or over an FSB for a CPU.

Some cards cannot do high-precision FPU operations- their shaders and such will only do 32-bit math and can't handle double-precision FP calculations. I know that AMD's R600 generation can handle 128-bit calculations and I know that NVIDIA's G70 can only do 32-bit math, but I'm not sure about the capabilities of the G80 and G90 series or any of Intel's IGPs.

So to sum it up, GPUs are wonderful for things that have a relatively small working set and require lots of memory I/O. Other things are better suited to a CPU, particularly operations that need to access a very large amount of memory. If you want to ask an expert on this, I'd ask the guys over at the Folding@Home forums as FAH is an HPC app that runs on GPUs as well as CPUs.

 

[Shadow703793]

^True.

 

[dobby]

i hope everyone is folding at home? i know i am. (im not in the THG team though-sry)

 

[Shadow703793]

I used to FAH but lost touch of it after a while.

 

[Amiga500]

Why don't I throw in CUDA and a 9800GX2 for the Xeon/Itanium mix ? . The GPUs can actually calculate Floating point and smiler operations faster than a CPU and cost less.

I've been asking about this being ported to CFD work.


Unfortunately, memory limitations, instruction limitations and a vested interest in not having software run too quick without parallel licenses present obstacles.


When fusion comes along, if the GP-GPU has access to the main memory at a low latency (and high bandwidth) then the first obstacle is removed.

Inclusion of a few more instructions in the pipelines might fix the second problem. Essentially turn the GPU into a RISC processor.

 

[cyberunlimited]

Itanium hardware is a lot better because its not bag down by the architectural requirement of X86.
Apples and Oranges...

Itanium is more of RISC(Reduce Instruction Set) type CPU and X86 is a CISC(Complex Instruction Set Computing)

Itanium can emulate x86 through software but it will be slower. Thats the reason Intel can't make Itanium x86 compatible. Intel will have to add more tansistors/circuitry to make Intanium x86 compatible.

RISC Type Processors:

Power PC (IBM , Freescale, PS3-Cel CPU designed by IBM)
SUN SPARC Processors

CISC TYPE Processors:
Xeon x86
Pentium Series X86
AMD X86
MCS 51/96 Micro-controllers

RISC Advantage
1) have fewer instruction sets
2) Run at a higher clock rate
3) Lower Transistor Count
4) Higher efficiency.... Multiple Instructions Execution per Clock Cycle.
5) Ist not bug down by the DEEP-INTRUCTION PIPELINE needed by CISC such as X86 Architecture.


Big Iron Machines for Scientific and design applications mostly use RISC. The LOS_ALAMOS labs that simulate nuclear explosion. Circuit Simulation & Synthesis that takes months to compute normally use RISC. These supercomputers have > 10K CPUs working together. Common software is UNIX/Linux Operating system, Design , Mathematical and Engineering softwares

CISC is mostly for office application, business, banking, games. CISC servers are cheap than RISC thats why its very popular and widely use. And of course MS-Windows, and MS-Office is probably 85% of the main application.

Don't forget one of the best 64Bit RISC Operating System every; OpenVMS.

 

[Mousemonkey]

This topic has been closed by Mousemonkey