Itanium 2 has begun shipping

[Athlon4all]

Here's Intella's Press release

 

[dullard]

Well I'm glad I was right about the processor prices, and Intel didn't increase them. Now it is sit back and wait time. I truely think Itanium has the capability of being quite big - if only Intel can convince customers to use it. Or it could be a great opportunity turned into a flop. I'd really like to see the expensive proprietary processors go the way of the dinosaurs. Not that I want Sun, SGI, etc to go out of business - I just think their resources can go to a better use if they switch to Hammer and Itanium. Plus it will save the consumers millions (billions?) of dollars.

It is obvious that Intel expects this to catch on slowly. Several places in that link had the phrase "deliver...Itanium 2-based systems...over the next year". Meaning don't expect to see it selling in any significant quantities right away. Plus with them advertizing a better chipset that won't be available till fall, expect zero sales until then.

Just for my curiosity, I'd like to see Intel release 1 MB cache versions for high-end workstations. That would probably push the cost to just under $1000 per processor. Many workstation programs don't benefit from that much cache.

 

[pm]

It is obvious that Intel expects this to catch on slowly. Several places in that link had the phrase "deliver...Itanium 2-based systems...over the next year". Meaning don't expect to see it selling in any significant quantities right away. Plus with them advertizing a better chipset that won't be available till fall, expect zero sales until then.

That's Intel's chipset. There are other chipsets from other vendors out there already or coming in the near future. For example, HP is currently shipping with their own zx1 chipset.

Just for my curiosity, I'd like to see Intel release 1 MB cache versions for high-end workstations. That would probably push the cost to just under $1000 per processor. Many workstation programs don't benefit from that much cache.

The 1.5MB and 3MB cache versions both use the same die. Making a 1MB version wouldn't save any money because the die size is the same and would introduce some interesting issues due to the way the L3 is set up.

 

[dullard]

That's Intel's chipset. There are other chipsets from other vendors out there already or coming in the near future.

That statement was from the viewpoint of a reluctant customer. Many business people have a viewpoint that an Intel chipset is best for Intel chips. With Intel devoting so much space in the press release about their great new upcoming chipset - that only keeps that idea strong. Thus my conclusion is that a significant number people considering Itanium 2 will wait to see the Intel chipset before making a purchase.

The 1.5MB and 3MB cache versions both use the same die. Making a 1MB version wouldn't save any money because the die size is the same and would introduce some interesting issues due to the way the L3 is set up.

That comment was just a hypothetical situation. I didn't know there would be L3 issues (its nice to have an Intel employee around that is willing to educate us on advanced info like this). But hypothetically Intel could reduce the die size for the lower cached versions if those issues could be resolved. Maybe it isn't possible the way Itanium 2 was designed, that is fine - I was just thinking out loud that it would have been nice to see lower priced low cache versions.

I just think Intel had more cache than necessary if it wants to be a high-end workstation product. When the original Itanium was released at 4 MB cache, the typical workstation processor had only 256 KB cache. That huge cache does add to the price - helping pushing it out of the workstation market. Intel reduced the cache from 2/4 MB down to 1.5/3 MB which is better for many workstation needs. I realize that servers might need the high cache - so Intel doesn't want to lower it too much. But really 1.5 MB is just more than needed for workstations in my opinion.

Edited - this post seemed combative but that wasn't my intent. Sorry.

 

[pm]

That statement was from the viewpoint of a reluctant customer. Many business people have a viewpoint that an Intel chipset is best for Intel chips. With Intel devoting so much space in the press release about their great new upcoming chipset - that only keeps that idea strong. Thus my conclusion is that a significant number people considering Itanium 2 will wait to see the Intel chipset before making a purchase.

Companies like HP (and the other powerhouses) see chipsets as their way to "value-add" and differentiate themselves from other companies. Large "iron" companies have a substantial reputation of reliability and uptime for their chipsets. I would say that Intel is the "new kid on the block" in the server and workstation space. Which is not to trivialize the impact of the Intel chipset, but I don't think that people in the high-end computing space pay as much attention to the chipset vendor as they do the performance, features, price and reputation of the computing vendor.

Intel wised up and reduced the cache from 2/4 MB down to a better 1.5/3 MB. I realize that servers might need the high cache - so Intel doesn't want to lower it too much. But really 1.5 MB is just rediculous for workstations at the moment. It is possible that Itanium 2 has an identity crisis - customers might wonder is it a server chip or a high-end workstation chip? If they don't know right away, they will not buy Itanium.

The HP C3700 workstation (link to the HP page with the workstation overview) that I am typing this on right now has a 2.25MB L1 cache, and I've been using this workstation for well over a year now. Most of the applications that I run are in the 1-2GB range and take most of a day to run. From my perspective as an engineer, it seems to me that ~1-2MB caches are a drop in the bucket, but the more the better. As far reducing the cache from Itanium to Itanium 2, the reduction in the cache sizes has less to do with "wising up" as it does that the 2/4MB cache of the Itanium was off-die, and when the cache was moved on-chip in the Itanium 2 there was simply not enough space to put 4MB on-die.

 

[dullard]

PM. I'm in the computational fluid dynamics field. I've had the pleasure to do simulations on proprietary machines by Sun, SGI, HP as well as normal Intel workstations. Most computer purchases I've seen are in the $10,000 - $20,000 range (although some I think were much more costly). Itanium 2 has great potential with its record breaking floating point scores that Intel estimated a while back. However there needs to be a low cost Itanium 2 chip to fit a nice multiprocessor system into the $10,000 - $20,000 price range before I'd seriously consider it. ($30,000-$40,000 is possible if it really performed well). I know there are many people like me waiting to see how the Itanium 2 pans out and if it will fit in a range similar to mine.

In my experiences (I admit I've only done this for 4 years though), extra cache doesn't improve the speed for CFD calculations. My calculations tend to run from one end of memory to the other without repeatedly accessing the same data. Cache speeds things up when data is repeatedly used. Thus if your program doesn't repeatedly use data, then cache doesn't do much. The benchmarks I and others have done have shown that in every case where cache was doubled, performance was unaffected. So to me, "the more the better" doesn't apply.

Basically I'm trying to point out some things that Intel needs to consider (or probably already has). I doubt I can get an eight processor 900 MHz Itanium 2 for $20,000 - but if that were possible it would open up a whole new batch of customers who are currenly buying small clusters instead. Any cost cutting that Intel can do would inch the Itanium 2 that much closer to success (and my idea of a low cache Itanium 2 was one solution that I thought might work). Typicaly rule-of-thumb is more cache = larger die = higher prices (which PM graciously pointed out isn't the case with Itanium 2). I'd love a low cost Itanium 2. Comparing the prices of the 2MB to 4MB cache Itaniums and the 1.5MB to 3MB cache Itanium 2s the rule of thumb appeared to still true. I was unaware that Itanium 2 couldn't cut any more cache to cut the prices further.

 

[Athlon4all]

I have one question. What has Intel Changed for Itanium 2 besides the L3 Cache size? They increased the fsb bandwidth (according to the press release to 6.4GBs, and did I read the pr correctly, it is a point to point fsb like the EV6 is?), but otherwise, what makes Itanium 2, Itanium 2? The L3 Cache is still off die correct but it runs at the same clock speed as the CPU correct? I am only curious. I suppose I will find out in a few days when Chris over art Ace's Hardware posts part 3 of the Volume MP Systems which will cover the Itanium 2.

Oh pm/dullard, I am just wondering, in Ace's hardware's recent Opteron and Itanium Editoral, they said that in their opinion Many workstation users believe that more than 4Gig of Memory isnt enough, I'm just wondering what your opinion is Thanx

 

[pm]

I have one question. What has Intel Changed for Itanium 2 besides the L3 Cache size? They increased the fsb bandwidth (according to the press release to 6.4GBs, and did I read the pr correctly, it is a point to point fsb like the EV6 is?), but otherwise, what makes Itanium 2, Itanium 2? The L3 Cache is still off die correct but it runs at the same clock speed as the CPU correct? I am only curious. I suppose I will find out in a few days when Chris over art Ace's Hardware posts part 3 of the Volume MP Systems which will cover the Itanium 2.

The Itanium 2 is a completely different microarchitecture from the first Itanium. The difference between the two is similar to the difference between the Pentium III and the Pentium 4 (or any other new implementation of the same instruction set architecture). The pipeline is shorter, it clocks faster (don't ask why. ), the integer and FP execution unit are wider, bypassing is faster in the integer unit, the caches are bigger and the latencies are lower (again, don't ask why ), the FSB is twice as wide and 3MB of cache is fit on die and it's on the same process technology (0.18um). Fundamentally there is very little that is the same between the Itanium and the Itanium 2 except that they run the same instruction set.

Oh pm/dullard, I am just wondering, in Ace's hardware's recent Opteron and Itanium Editoral, they said that in their opinion Many workstation users believe that more than 4Gig of Memory isnt enough, I'm just wondering what your opinion is Thanx

I have 4GB on this machine. I would only need more if I had more CPU's to run with since fundamentally my job is computationally constrained, not memory constrained. (ie. I could get more done faster if I had more computational power, but more memory wouldn't help much). I can't see that more than 4GB would help much on this machine - running more jobs in parallel (which would require the additional memory) would just slow everything down, and I pretty much never run more than a 1.5GB thread - although occasionally I'll kick off two. Still, that's only 3GB. Looking forward, I can see that more memory will be required. But for now 4GB is more than enough for what I do.

 

[dullard]

Originally posted by: Athlon4all
I have one question. What has Intel Changed for Itanium 2 besides the L3 Cache size? They increased the fsb bandwidth (according to the press release to 6.4GBs, and did I read the pr correctly, it is a point to point fsb like the EV6 is?), but otherwise, what makes Itanium 2, Itanium 2? The L3 Cache is still off die correct but it runs at the same clock speed as the CPU correct? I am only curious. I suppose I will find out in a few days when Chris over art Ace's Hardware posts part 3 of the Volume MP Systems which will cover the Itanium 2.

Oh pm/dullard, I am just wondering, in Ace's hardware's recent Opteron and Itanium Editoral, they said that in their opinion Many workstation users believe that more than 4Gig of Memory isnt enough, I'm just wondering what your opinion is Thanx

The L3 is on-chip for the Itanium 2. Plus there is the marginal clock speed boost and optimizations. As for other changes, we will wait for PM to respond.

I actually posted my responce to your memory question
here. But I'll be glad to go into more details about my experiences. A model of turbulent, three dimensional, time dependant flow solving for 7 million unknown variables will take roughly 3 weeks to finish on a dual 1.7 GHz Xeon (and it will use right around 1 GB of memory). If I used the same computer to run a simulation requiring 28 million unknowns, it will need roughly 4 GB of memory. However it will now take roughly 3 months to complete on the same computer. I think anyone would agree that that is unacceptable. Even a dual 4 GHz machine will take 1 month or so (maximum speed that I'd expect to see before 2004). Still far too long for a simulation. Instead if I need a simulation that is that complex, I'd spread it out on 10 or 20 processors in a cluster, each computer in the cluster needing only 512 MB of memory (except the main computer in the cluster which will need a bit more). So for me the 4 GB barrier is not important for the next few years.

That article also mentioned CAD users switching from wireframe to fully rendered pictures. That caused a need for more memory - often up to 2 GB for a great CAD machine. That however was a one time jump. There is nothing else on the horizon that will suddenly need to push a CAD user's need past 2 GB.

At the moment the best selling workstations have a 2 GB limit - meaning that most workstation users don't have a current need for more than 2 GB. I just cannot forsee all these users suddenly needing a memory jump to 4 GB by 2004. Memory use has started to level off. Average computer memory used to almost double every year. This growth has slowed to about a 35% increase every year. Using the 35% estimate a computer user needing 2 GB now will still only need 3.6 GB in mid 2004. I think that article you mention ignores this leveling off effect and assumes the doubling each year will continue.

Edit: PM beat me to it. We are both CPU limited not memory limited. By 2005 that will change - but hopefully 64 bit will become available for desktops by then.

 

[Jeff7]

Itanium 2?? I didn't enven know that they had even really begun large sales of the original Itanium. Have I just been totally out of it or something?

 

[dullard]

Originally posted by: Jeff7
Itanium 2?? I didn't enven know that they had even really begun large sales of the original Itanium. Have I just been totally out of it or something?

Intel is on a really fast pace with Itaniums out every summer. Blink and it is easy to miss a release.

 

[Athlon4all]

Thanks pm/dullard so it really depends about the memory. Anyway, I understand now about the Itanium 2, cool Thanx again

 

[pm]

Here's a list of papers and presentations on the Itanium 2 ( I co-authored the "The 16kB Single-Cycle Read Access Cache" paper). The presentations tend to be more readable but YMMV. If you are comparing Itanium to Itanium 2, slide #7 of the the IDF presentation of the Itanium 2 shows the differences graphics (Note: .PDF format). In addition to the IDF presentation, there's a pretty interesting overview of the Itanium 2 here. (link to ISSCC 2002 presentation of McKinley in .PDF format. Note that ISSCC is a circuit design conference, so most of it's focused on circuits, but the first set of slides is more general information with some neat die photo shots).

 

[grant2]

PM do you have any thoughts/speculation you can share on intel's 64-bit computing plans? Assuming that 64-bit will eventually make its way all the way down to home PCs, do you think intel will try to drive itanium's instruction set as the only 64-bit solution? Do you think they'll attempt to create 64-bit extensions for IA32 like AMD has done? Or do you think they'll (gasp) adopt AMD's extensions?

 

[Athlon4all]

heh Grant, you're putting pm on the spot, trying to get him to tell you Intel's confidential plans. My personal opinion is that for certain, Intel will not adopt both x86-64 and IA-64. Its either or, and I am leaning towards IA-64 simply because 1. all the money that has been spent on it by Intel and HP 2. This is the time for x86 to stop and be replaced, now is EPIC the best solution, I have no clue, but it is getting to be time to get rid of x86. We will see. Most likely I feel is that Intelmay for a few cpu cores make a Dual Core of some sort (likely on .09um) that will provide equal x86 performance to a Pentoum 4 but support EPIC as well. Where does AMD fit in, good question

 

[Sukhoi]

There is a rumor going around (of which I'm sure pm can't comment on) that Intel is working on a P4 type processor with 64-bit processing if I remember right.

 

[formulav8]

If you go over to Ace's you will see that it is the typical Intel paper launch. Doesn't expect shipments, especially in high volume until 4th 1/4.

 

[pm]

grant2: As an Intel engineer, I'm always careful to avoid discussing future products. I know our roadmap out as far it exists today, but discussing future products outside the company is generally frowned upon (read: is grounds for being fired) so unfortunately I'm not able to discuss any aspects of it.

Having said this longwinded disclaimer, I know that Paul Otellini, our COO, has bluntly stated on several occassions that there will be no x86-64 for Intel or anything similar to it. Intel is completely commited to IA64 and currently there are over 5 IA64 future products currently under development within my division. The Itanium 2 is, in my mind, a very compelling product. It has the performance it needs at a competitive price point. With the two biggest companies in the computing business (HP/Compaq and IBM) solidly behind it, the question is not whether it will be a success but merely how long will it take. From my perspective, the only thing truly holding it back at this stage is software, but those who write software have commited to supporting it so it's merely a matter of time. I know that they are starting to beta test Itanium 2 workstations running 64-bit Linux within my team to replace our current HP C3700 workstations running HP-UX and everyone involved has had only good things to say about the switchover. The performance is more than double on the apps that I run and this takes something that runs 10 hours and turns it into something that only 4-5. So instead of waiting 2 days for a run to complete, we can run something in the morning, have results by the afternoon and kick off another run before we go home.

I personally am impressed with the preliminary plans for AMD's Opteron. The presentations that they have given make for interesting reading. It will be very interesting to see how this plays out. But there are a lot of risks involved in their strategy as well, so it's really too early to tell. We shall see.

 

[pm]

If you go over to Ace's you will see that it is the typical Intel paper launch. Doesn't expect shipments, especially in high volume until 4th 1/4.

First this is not true. Parts are shipping today. Intel is selling and shippping parts to OEMs. As far buying a system today, if you placed an order today you probably couldn't get one for quite a while because vendors are filling back orders. But Itanium 2 systems are shipping out the door today as well. So it's not a paper launch of CPU's or of systems. I worked on Itanium 2 for a long time and I'm still in contact with the team. They are shipping parts. I know this for a fact. It is not a paper launch.

Second, I went to Aces and didn't see this anywhere in their News section. Where are you getting this information from?

 

[RaynorWolfcastle]

Originally posted by: pm
grant2: As an Intel engineer, I'm always careful to avoid discussing future products. I know our roadmap out as far it exists today, but discussing future products outside the company is generally frowned upon (read: is grounds for being fired) so unfortunately I'm not able to discuss any aspects of it.

Translation We have killer products coming down the pipe unfortunately I can't discuss them yet

pm, are you still working on McKinley right now (fixing errata and such) or have you moved on to another project? (out of curiousity)

-Ice

 

[paralazarguer]

LINKY

This article states that the Itanium 2 now only uses an 8 stage pipeline with all sorts of crazy efficencies. That's much more efficient than the Athlon at the same clock frequency! Who says that AMD has the best IPC? Of course, it sucks at 32bit. It's only running at a ghz but it's still at .18u so who knows how far this architecture will go? I remeber when the new cores came out way back and people dissed them because they ran so poorly at 16bit code which is what everyone used (windows 3.1) The only thing that was capable of harnessing the power of these crazy 32bit CPUs was UNIX but no one cared because everything was windows 3.1. I remember all the magazines criticizing the need to move to 32bit etc and calling the new architecture inefficient. I can't imaging going back to windows 3.1. My God, what were those "experts" thinking. I really think it's the same thing now.

 

[pm]

Translation We have killer products coming down the pipe unfortunately I can't discuss them yet

That's one way to translate it. But what am I supposed to say? It's the truth. And I like my job.

pm, are you still working on McKinley right now (fixing errata and such) or have you moved on to another project? (out of curiousity)

I'm on another project which has yet to be publicly revealed. I stopped working on McKinley about a month ago. I would have finished a lot sooner, but I worked on some of the back-end validation stuff (I was the technical lead for the team developing the manufacturing tests that check to make sure all of the transistors on all of the chips turn off and on correctly). And, completely honestly, I never had to fix any errata. The errata fairy never blessed me with a single problem in any of my units and I was responsible for the circuit design and mask layout of about half a million transistors during McKinley. A little bit of luck (plus I'm pretty conservative as a circuit designer. I mostly use static CMOS and don't take a lot of risks... Not very exciting, but at least I don't get phone calls in the middle of the night to come over to the lab to look at signal waveforms in my units).

 

[grant2]

heh Grant, you're putting pm on the spot, trying to get him to tell you Intel's confidential plans.

I know you're joking but I used the words "can you" on purpose because I don't want him on the spot...

This is the time for x86 to stop and be replaced, now is EPIC the best solution, I have no clue, but it is getting to be time to get rid of x86.

"why fix what ain't broke?" When Macs changed architectures, many years ago, all the zealots thought apple had wintel beat because "x86 doesn't have any more room to grow" Where are we at now? Only photoshop fanatics would deny that top-tier x86 cpus have not only kept up with, but outpaced apple's "new" architecture.

The "time to get rid of x86" changes totally depending on who you talk to. If you ask a pure engineer, the time to get rid of ANYTHING is 5 minutes after it's superseded- even if it means starting over from scratch every few months. If you ask the average user/business, the time to get rid of something is when it can no longer perform it's function. And can x86 continue performing its function? You bet it can, especially when extended to 64bit.

Maybe i'm a bit of an anomoly in a technical forum because I don't derive any particular pleasure from knowing i'm using an "elegant" instruction set... but I know 99% of the general population feels similarly.

is EPIC the best solution, I have no clue,

Personally, I think Transmeta has the best way to achieve and maintain an elegant instruction set. If it lives up to its promise, you can enjoy all the benefits of constantly improving the instruction set without any of the pain of migrating code or emulating older sets.

 

[pm]

Personally, I think Transmeta has the best way to achieve and maintain an elegant instruction set. If it lives up to its promise, you can enjoy all the benefits of constantly improving the instruction set without any of the pain of migrating code or emulating older sets.

I'm not sure that I agree. This method adds another abstraction layer to the CPU. It adds flexability but at the cost of performance. The words "if it lives up to it's promise" are interesting because the truth is that the hype that preceeded Crusoe could never have been met. The fact is that translation can never be perfect and there is always a penalty. A well-thought out hardware implementation will always be faster. There are places where translation makes sense - but only if you are willing to knowingly sacrifice performance.

Transmeta is far from being the first company to implement binary translation on-the-fly. They are merely the first that I have heard of who made it their core business model and to implement so many hardware hooks to support it. But binary translations are used by several companies - Compaq has a binary translator on their Alpha processors (FX!32). Sun has used it (Wabi). HP is using it (Dynamo). IBM has used it. It makes sense in specific situations - mainly ones in which you don't want to waste the space/time/effort/money to implement the functions in hardware.

 

[andreasl]

Hey pm is there any info about Itanium 2 32-bit performance? I read somewhere that it was supposed to be vastly improved over Itanium 1. Obviously not to the degree that it will compete with XP's or P4's but I would be interested to know if it was improved and by how much.