How 'slow' is slow for 32bit on Itanium?

[AgaBoogaBoo]

I'm just wondering how slow actually running a program in 32bit mode on an Itanium is compared to running it in 64bit compiled program. Also, how slow is running a program on the Athlon 64 in 32bit mode compared to running the same program in a 64bit compiled program?

I'm not really interested in buying either of these, but I'm curious about it because people always say that the Itanium has slow 32bit processing and others say that in a 64bit server, there isn't a need to run 32bit programs very fast.

 

[AtomicAlien]

The Athlon 64 wouldnt be affected since it can process 32-Bit code in it's native form anyways, so no problem there.

The Itanium is a different story, I heard that a 733MHz Itanium running 32-Bit code is as slow as a 700Mhz Pentium3, which for modern 32-Bit server CPUs is horrible. Intel is supposed to be releasing a new firmware for the Itanium to help improve this...

Just my two cents!

 

[Lord Evermore]

We don't know yet exactly how well an Athlon64 runs 32-bit code, and can't know how well it runs 64-bit code yet because there are few or no applications supporting x86-64, and Linux is the only OS that you might be able to get that runs in 64-bit mode on it right now. The various Opteron reviews can give you an idea of how the Hammer architecture does with existing 32-bit code. Basically Athlon64 shouldn't run much different from an AthlonXP on 32-bit code, and hopefully a good bit faster due to the optimizations, but as with the P4, the amount of work done per clock is actually less, so it could be a bit lower performance as well, but able to scale to higher speeds. As far as 32-bit code though, the A64 is just like if they'd designed the Athlon with the same pipeline length and L2 cache, and with an integrated memory controller. It runs completely as a 32-bit CPU if the OS is only 32-bit.

If the OS is 64-bit, then an application can run in 32-bit mode with no performance hit, and you can also run 64-bit code. The biggest thing for 64-bit is the memory addressability, but all the extra registers will probably result in better performance compared to the same application in 32-bit mode.

Itanium on the other hand can't natively run 32-bit code at all, or even any sort of x86 code. It has to translate x86 instructions to the IA-64 instruction set before it can work on the instructions. Severe performance hit. This was why we expected Intel to put a complete 32-bit core on the die with the Itanium core, but they changed their mind for some reason.

 

[AndyHui]

Running regular 32-bit x86 code, the original Merced Itanium had the equivalent performance of a Pentium 166.

 

[AgaBoogaBoo]

Ouch, hopefully that has changed with Itanium 2.

 

[SexyK]

IIRC, the new software for Itanium should allow 32-bit code to run around the same speed as on a 1.5GHz P4. Noone buys an Itanium to run 32-bit code though.

 

[LikeLinus]

Originally posted by: SexyK
IIRC, the new software for Itanium should allow 32-bit code to run around the same speed as on a 1.5GHz P4. Noone buys an Itanium to run 32-bit code though.

Thats what I was going to say. If you have the cash to spend on an Itanium, I hope you'd be smart enough to know better than running 32bit apps.

People who use 64bit procs don't use 32bit apps.

 

[Pariah]

Originally posted by: AgaBooga
Ouch, hopefully that has changed with Itanium 2.

Why? Anyone buying an Itanium system to run 32bit code should be fired. Why would you have a high dollar system like that running 32bit emulation when for $400 you can get a 2.2GHz P4 Dell with monitor that will run it magnitudes faster?

 

[dullard]

Thats what I was going to say. If you have the cash to spend on an Itanium, I hope you'd be smart enough to know better than running 32bit apps.

People who use 64bit procs don't use 32bit apps.

I can just see it now:
"Excuse me, Boss, can I interupt the server to write my report on, this desktop of mine just isn't fun enough with Word..."

I personally haven't used an Itanium server. But I have used plenty of other ones (SGI, Sun, HP) to run complex simulations. In each case, we do the complex work on the server, then shift the work over to cheap workstations to do the post processing. That way someone else can use the expensive server. Everyone else I know at several universities does the same thing. Any company with enough resources for an Itanium server probably has enough for a desktop for their 32-bit work.

 

[AgaBoogaBoo]

Originally posted by: dullard

Thats what I was going to say. If you have the cash to spend on an Itanium, I hope you'd be smart enough to know better than running 32bit apps.

People who use 64bit procs don't use 32bit apps.

I can just see it now:
"Excuse me, Boss, can I interupt the server to write my report on, this desktop of mine just isn't fun enough with Word..."

I personally haven't used an Itanium server. But I have used plenty of other ones (SGI, Sun, HP) to run complex simulations. In each case, we do the complex work on the server, then shift the work over to cheap workstations to do the post processing. That way someone else can use the expensive server. Everyone else I know at several universities does the same thing. Any company with enough resources for an Itanium server probably has enough for a desktop for their 32-bit work.

Thats what I was thinking as well.

 

[pm]

Running regular 32-bit x86 code, the original Merced Itanium had the equivalent performance of a Pentium 166.

Extrapolating the Itanium's ability to run 32-bit code from that article in which they installed a 32-bit version of Windows and bencharked it is not very realistic. Running 32-bit Windows with 32-bit applications results extremely slow performance. With either the Itanium or the Itanium 2, running 32-bit apps on a 64-bit version of Windows produces much better results.

Itanium on the other hand can't natively run 32-bit code at all, or even any sort of x86 code. It has to translate x86 instructions to the IA-64 instruction set before it can work on the instructions. Severe performance hit. This was why we expected Intel to put a complete 32-bit core on the die with the Itanium core, but they changed their mind for some reason.

When you say "natively run 32-bit code" what do you mean? It has a hardware instruction decoder that translates IA32 instructions into micro-ops which are then dispacted to the execution pipes. How is this different than the way the Althon, the Pentium III and the Pentium 4 operate?

As far as the original question, I honestly don't know the answer, although I suppose that I should. I haven't benchmarked IA32 code on an Itanium or Itanium 2. We have been using Itanium 2 workstations, and my general useage pattern is to use IPF binaries for high-performance apps, and use the IA32 binaries when I can't find a IPF binary or I can't recompile it.

 

[Lord Evermore]

By native I just mean that it's not designed to be an x86 processor, which the others are, and doing x86 code is just sort of tacked onto it. Also because it's a completely different type of architecture (EPIC) than anything previous. Athlon or P4 may be breaking it down into micro-ops, but not having to change it as much as Itanium.

Intel should switch to all Itanium machines, running 32-bit OSes. One: it'll punish them for all their wrongs, and two: it'll slow down the damn megahertz race, since designing a processor would take 16 times as long.

 

[AtomicAlien]

Originally posted by: AndyHui
Running regular 32-bit x86 code, the original Merced Itanium had the equivalent performance of a Pentium 166.

Yow. That's even worse than I thought...

 

[AndyHui]

Extrapolating the Itanium's ability to run 32-bit code from that article in which they installed a 32-bit version of Windows and bencharked it is not very realistic

I know. To me, I think people are having a hard time understanding that you just are not supposed to run 32-bit code on the Itanium.

 

[HokieESM]

Originally posted by: dullard

Thats what I was going to say. If you have the cash to spend on an Itanium, I hope you'd be smart enough to know better than running 32bit apps.

People who use 64bit procs don't use 32bit apps.

I can just see it now:
"Excuse me, Boss, can I interupt the server to write my report on, this desktop of mine just isn't fun enough with Word..."

I personally haven't used an Itanium server. But I have used plenty of other ones (SGI, Sun, HP) to run complex simulations. In each case, we do the complex work on the server, then shift the work over to cheap workstations to do the post processing. That way someone else can use the expensive server. Everyone else I know at several universities does the same thing. Any company with enough resources for an Itanium server probably has enough for a desktop for their 32-bit work.

Yep.... that's mostly how it goes here at VT (I'm in grad school doing computational solid mechanics). We use SGIs and/or HPs (some with Itaniums, some old ones with Alphas or R12000s) for our modeling.... and they're all 64bit (which isn't hard, because we're writing the source code--so we just need a 64 bit compiler). We use plain-old Dell workstations (P3s and P4s) for post-processing.

Of course, what's shocking is how fast a P4 (or Athlon XP) can be, even for numerical simulations, if the code is written well for it (like not requiring more than 2GB of memory, etc).

That said, from what I understand (from more informed people than I), that the Athlon64 should run 32 bit code about the same as a similarly CLOCKED (not PR, the actual MHz) Athlon XP. There might be a small advantage (with extra memory registers).... but that's probably a good comparison. Which isn't much to write home about (the Opteron is up to what, 1.8 GHz).... but that's not including the 64-bit executables. As soon as a reliable 64-bit OS (and a reliabe 64-bit C compiler) is around.... THEN we'll be talking (double precision floating point really really benefits from 64bit).