Great 64 bit Processor Article

[zackbass]

I was making my daily rounds of the internet and came across a gem of an article on ExtremeTech about the new 64-bit processors. It'll take you a while, but is very interesting in the comparion of the Hammer and IA-64 processors. After reading what they have to say, I dare announce that my next processor will be a 64-bit Intel.

[yiotis]

I apologise if i disappoint you but we shall not see 64-bit intel cpus soon.I think this will happen after the pentium 4

edit:i am always talking for desktop computers

[zackbass]

Originally posted by yiotis
I apologise if i disappoint you but we shall not see 64-bit intel cpus soon.I think this will happen after the pentium 4

edit:i am always talking for desktop computers

I suppose you didn't take into account that I can't afford a new processor until about the time these come out .

[Moridin]

Originally posted by zackbass
I was making my daily rounds of the internet and came across a gem of an article on ExtremeTech about the new 64-bit processors. It'll take you a while, but is very interesting in the comparion of the Hammer and IA-64 processors. After reading what they have to say, I dare announce that my next processor will be a 64-bit Intel.

IA-64 isn't intended for use in Desktop PC's, it's designed for large servers and high end workstations.

I apologise if i disappoint you but we shall not see 64-bit intel cpus soon.I think this will happen after the pentium 4

The P4 trace cache could make it very easy to add 64-bit support to the P4. It isn't out of the question that 64-bit operations could be converted to 32-bit ops for internal execution. There would still need to be some support for 64-bit registers though...

[Un4given]

But did you read the article carefully? IA-64 will rely on the software and the compiler used to create that software, to perform efficiently. Do you know how hard it is to write a compiler? It's not an easy task, and may take quite some time to develop the compiler to the point where it actually works well enough for the processor to show it's power.

If Intel wants IA-64 to be truly and quickly successful, in my opinion, they need to make the CPU capable of executing x86 code without taking a performance hit compared to the CPU that they just replaced. People aren't going to be really keen on replacing every piece of software they own to support native IA-64 language, or take a hit in performance compared to their previous CPU when executing their existing software.

[zackbass]

I'm sorry for not specifying that most of the things I do on my fast computer are video and photo editing, and I hope to get started with some CAD and programmming. I'd rather be out in the garage on the lathe/milling machine, or riding my mountain bike than playing games and other desktop apps at home (though I sometimes do). I'd like to buld a "real" high-end workstation/part-time server (for the rest of the boxen) once I save enough cash, which I will have plenty of time to do.

BTW: I wasn't trying to start something big by pointing out the article, just pointing out an informative read.

[Moridin]

Originally posted by zackbass
I'm sorry for not specifying that most of the things I do on my fast computer are video and photo editing, and I hope to get started with some CAD and programmming. I'd rather be out in the garage on the lathe/milling machine, or riding my mountain bike than playing games and other desktop apps at home (though I sometimes do). I'd like to buld a "real" high-end workstation/part-time server (for the rest of the boxen) once I save enough cash, which I will have plenty of time to do.

BTW: I wasn't trying to start something big by pointing out the article, just pointing out an informative read.

Sounds reasonable, it’s just that there are a lot of misconceptions out there about what this processor is for. It’s designed to compete with RISC processors running commercial Unix like Solaris, AIX, Tru64 or HP-UX, not Windows desktops.

On native applications it is FP monster and I don’t think anything will touch it on technical computing or graphics rendering when it uses native applications. On the server side performance is usually more system/chipset related then anything else. I don’t think Itanium would make much sense in a small server.

[Ramuman]

Originally posted by Un4given
If Intel wants IA-64 to be truly and quickly successful, in my opinion, they need to make the CPU capable of executing x86 code without taking a performance hit compared to the CPU that they just replaced. People aren't going to be really keen on replacing every piece of software they own to support native IA-64 language, or take a hit in performance compared to their previous CPU when executing their existing software.

Putting full hardware x86 support would be just about the most counter-intuitive thing Intel could do. When you say the CPU that they just replaced...what do you mean by that? IA-64 is entirely different line of high end CPUs that Intel had no predecessor to. Xeon was probably the closest and it was a mile away from what IA-64 is. People that buy IA-64 system will not care about replacing existing software. In fact, if Intel can get a few select programs over to IA-64, they'll have the software issue settled. AFAIK, they've already done this with Oracle, Unix etc., and that's really all they need.

IA-64 doesn't need a IA-64 version of AOL IM to succeed. It needs server applications and as far as that goes, Intel's done their fair share.

The fact of the matter is that moving to IA-64 wouldn't be that much more difficult then say getting any new line of servers and upgrading software. For example, say you operate a datacenter and your running a few key aps, namely Unix/Linux, Oracle, DB2 etc. If you need to update your servers, you would likely be looking into updating the software as well. What difference does it make if you go USIII, Power, Alpha or IA-64 if they all support the same or similar aps and your total cost of upgrade won't be that different or important between the choices? The answer then is simple, it comes down to if the IA-64 platform itself can compete with the others. Intel is a proven name, but even they have never been in a market this high up, and thats going to be the true test.

IA-64 in its current form will never become a desktop ISA. It can however contribute to elements of a desktop ISA in the extended future. x86 support was something Intel threw in there to maintain a continutity they haven't lost since their first IC 4004. If ones buying an Itanium or McKinley server for x86 applications, then they're doing it for all the wrong reasons and their IT group has the combined intelligence of a dead fish.

[Sledgehammerhead]

I'm sorry to say, but that is one of the most poorly written articles that I have ever seen. The comparisons that were drawn were almost 100% falicy, and the conclusions that were drawn were almost 100% questionable.

I'm not saying that Itanium is a sinking ship. I think it's a little early in the game to relegate either chip to the trashcan, but the author of this article makes extremely questionable statements all throughout the article. The conclusion page has some especially ugly gems:

Backing Intel's newest and heavily promoted next-generation architecture is a foregone conclusion for vendors that want to stay in business.

It's too bad that just about every OEM, including Dell has dropped Itanium from their shipping servers... It's too bad that all of these OEM's don't want to stay in business.

Both processors are totally, completely, and inarguably backward compatible with x86 binaries. Anything else would be a criminal dereliction of duty.

The Itanium is widely known to run x86 code at the speed of a 200 Mhz Pentium tops. It may be compatible with x86 binaries, but what good is that if the system cannot run modern x86 code faster than a Commodore 64???

The author consistently refers to x86 as the noose that is about to hang processors from both Intel and AMD. While x86 is not elegant, it has been proven that x86 processors are massively scalable and massively fast at processing their code. Even the 10 stage Athlon has scaled to near 1.7 Ghz on current manufacturing processes, while the Itanium has been held back at less than 1/2 the clockspeed (800 Mhz). I realize that the Itanium is an IPC marvel, but frankly, so was the K6-3, and I rarely see people touting the merits of this processor.

IA-64 code will average about 3x the size of the equivalent x86 code. With code this size, cache size, memory size and bandwidth for both will be extremely important. Intel has taken this into consideration in their architecture design, but not sufficiently, IMO. The cache is excellent, but for a multi-Itanium system, the bus will very quickly become saturated, and the memory subsystem will be a huge bottleneck. On Hammer, on the other hand, memory performance *scales* with the number of processors, resulting in huge memory bandwidth available to the system. Athlon is a wonder of a processor, but has always been a little starved in filling its execution units. Having a larger, more efficient cache and on-board memory controller (hooked up to dual-channel DDR) will largely rectify this situation.

Both Hammer and Itanium have merits, but Hammer is a much better solution for the next 5 years in the desktop, workstation, and small to mid server space. Itanium may be decent for some big iron, but the costs of the processor and architecture will put it out of reach for 99.99% of America. So far, about 200 Itaniums have been sold, so that reaffirms my statement that it will be a niche product for a while. I don't expect to ever see a desktop IA-64 solution.

I've pointed out a couple of problems in the article, but to give a proper rebuttal for all of his errors would take a post that is 2x times the length of his article. Putting thought into word is a long process, and that is probably why his article is so short. Suffice it to say the actual information that is gathered from the article is nothing that hasn't already been told in white-papers for the past half-year. And the opinions that are so liberally strewn about the article just go to show that the author has no authority to be writing an article of this nature. I hope this post did not come out too abrasive, but basing your future buying decisions on this article is not a good idea.

--Sledge

[yiotis]

I have a question to pose to you guys just from curiosity:
Is there going to be a frequency gap between 64 bit and 32 bit cpus.
You see the Itanium runs at about 800Mhz and the P4 at 2000Mhz

[Sledgehammerhead]

I have a question to pose to you guys just from curiosity:
Is there going to be a frequency gap between 64 bit and 32 bit cpus.
You see the Itanium runs at about 800Mhz and the P4 at 2000Mhz

Yes. I would give a 95% chance of a continued frequency gap between P4 and Itanium and it's derivatives. There are a couple of reasons for this. P4 is a 20-stage pipeline, as we all know. Many of those stages are decoding micro-ops and pushing data around and doing things that don't do a lot of work but push the data around. The lengthy pipeline allows for some pretty crazy scaling. Itanium, on the other hand, is a 10-stage pipeline. I'm unaware about if they are going to be adding stages to McKinley (Itaniums next generation), but I do not believe this to be true.

IA-64 is an interesting instruction set in that the compilers do *a lot* of work. The amount of parallelism that the compiler can introduce will directly reflect on the performance of the system. But, IA-64 does not need as many "tricks" as x86 does to get performance out. Many of these "tricks" represent a stage of the pipeline in a processor, so due to the fact that these are not needed, I doubt a IA-64 processor will ever have a long pipeline.

IA-64 processors are also very low volume parts. As of yet, I do not believe that they are being produced on .13, like the NW's, and I doubt they are a very high priority to be moved over to .13 due to the present volumes. When the 300 mm fabs come on line for Intel and they have .13 silicon flowing out of their ears, they will move the production over, if IA-64 has any future at that time.

These processors, where they are sold, are also sold in areas where the consumers do not care what the clock speed is. They care about the performance under load and the performance of the specific system on the specific software that they are running. As the general market has been learning, clock-speed is nothing, and the big-iron folks have always known that.

I'm long winded... But the answer to your question is yes.

--Sledge

[Un4given]

Originally posted by Ramuman


Putting full hardware x86 support would be just about the most counter-intuitive thing Intel could do. When you say the CPU that they just replaced...what do you mean by that? IA-64 is entirely different line of high end CPUs that Intel had no predecessor to. Xeon was probably the closest and it was a mile away from what IA-64 is. People that buy IA-64 system will not care about replacing existing software. In fact, if Intel can get a few select programs over to IA-64, they'll have the software issue settled. AFAIK, they've already done this with Oracle, Unix etc., and that's really all they need.

IA-64 doesn't need a IA-64 version of AOL IM to succeed. It needs server applications and as far as that goes, Intel's done their fair share.

The fact of the matter is that moving to IA-64 wouldn't be that much more difficult then say getting any new line of servers and upgrading software. For example, say you operate a datacenter and your running a few key aps, namely Unix/Linux, Oracle, DB2 etc. If you need to update your servers, you would likely be looking into updating the software as well. What difference does it make if you go USIII, Power, Alpha or IA-64 if they all support the same or similar aps and your total cost of upgrade won't be that different or important between the choices? The answer then is simple, it comes down to if the IA-64 platform itself can compete with the others. Intel is a proven name, but even they have never been in a market this high up, and thats going to be the true test.

IA-64 in its current form will never become a desktop ISA. It can however contribute to elements of a desktop ISA in the extended future. x86 support was something Intel threw in there to maintain a continutity they haven't lost since their first IC 4004. If ones buying an Itanium or McKinley server for x86 applications, then they're doing it for all the wrong reasons and their IT group has the combined intelligence of a dead fish.

First of all, the article was dicussing x86-64 vs. IA-64 architecture. It wasn't meant as an Itanium vs. Hammer article, since these two CPUs aren't even competing for the same market. His references also heavily pointed to Windows based support, therefore lower end, including desktop applications.

As far as those switching platforms not caring about changing software too, that is totally untrue. These aren't $200 out of the box applications. These programs may cost thousands of dollars for the base install with additional charges for user licenses. There are also a great number of these companies using custom made software, for which they have spent vast sums of money, and are not going to just let go by the wayside because Intel isn't looking to provide decent support for backwards compatibility of x86 code.

[Sledgehammerhead]

I can see that I didn't quite answer the question that was asked. I answered if Itanium and its derivatives would remain at a lower clock than P4 and Athlon. The answer to this is yes.

I do not think it will be true for Hammer, however. There is nothing inherent to 64-bit processors that make it so that they will be slower clocked than 32-bit. It's just the way that most of them have been designed. Hammer is a 12-stage pipeline and it's execution engines are largely based off of Athlon. Having a longer pipe-line and a similar backend should allow it to actually scale higher than Athlon, but there is no reason to believe that it will have an IPC deficiency vs. Athlon. The branch prediction and on-board memory controller will help IPC and should more than negate the penalty for increased cost of missed branches (pipeline flush).

I hope that answers the question sufficiently.

-Sledge

[Ramuman]

Originally posted by Un4given


First of all, the article was dicussing x86-64 vs. IA-64 architecture. It wasn't meant as an Itanium vs. Hammer article, since these two CPUs aren't even competing for the same market. His references also heavily pointed to Windows based support, therefore lower end, including desktop applications.

As far as those switching platforms not caring about changing software too, that is totally untrue. These aren't $200 out of the box applications. These programs may cost thousands of dollars for the base install with additional charges for user licenses. There are also a great number of these companies using custom made software, for which they have spent vast sums of money, and are not going to just let go by the wayside because Intel isn't looking to provide decent support for backwards compatibility of x86 code.

I definitly agree with the cost of software not being cheap. A single CPU or user liscence of Oracle can run a grand easily. My point however was that noone is going to be upgrading from a legacy system like x86 (or any CISC system) to a RISC high end server without wanting new software. The same would hold true if a company out grew ANY low or mid range server and needed a true RISC mid to high end server. Also, the cost of software IS in these markets often overshadowed by the hardware and other factors like overall uptime and performance. Do you know how much a top tier USIII server or Power based server can run? Easily into the low millions for a single server with hardware alone and in the case of some very high end mainframes like the z/900, 10-15 million for hardware alone. At this point there are several important factors at play, and cost of software is a limited issue.

I do also agree that custom software could be expensive to port, but if said custom software was developed for the x86, then you would most likely have to replace it when moving to a RISC server. The fact that Intel put ANY x86 support in IA-64 itself could be considered a tab surprising. In fact NO top tier server CPU I know of supports the ISA that the company first established into the mainstream. For example, Power4 can't run 68xxx programs or 3xx series programs from 15 years ago. This isn't an issue to most if any buyer of a Power4 system.

Also, I did read the article, and the fact that he even considered Itanium (or IA-64) as a whole as a low end or Windows x86 CPU shows he at best as dissillusioned in what hes talking about. A single Itanium cartridge costs more then many PCs. Its not a cheap economical low end processor, but in the servers where it would be used, this again isn't an issue when the server might costs 10s or 100s of thousands of dollars.

[Ramuman]

Originally posted by yiotis
I have a question to pose to you guys just from curiosity:
Is there going to be a frequency gap between 64 bit and 32 bit cpus.
You see the Itanium runs at about 800Mhz and the P4 at 2000Mhz

I'm positive there will be a frequency gap. This should be a given. Almost no server CPU runs as fast clock speed wise as today's top-tier PC CPUs. They work very differently and this is a whole another topic, but to answer your question, IA-64 will likely never run as fast as a then current x86 CPU, but on that token, neither will a US, Power or Alpha. The Hammer is a different story and it should keep pace with other x86 CPUs (namely NetBurst core based ones), but again this is an entirely different CPU then IA-64 based CPUs.