Intel Itanium 2 Benchmarks

Pablo writes "Over at VR-Zone we saw some interesting benchmarks of the upcoming Intel Itanium 2 processor codenamed McKinley that is on schedule to be launched during second half of this year. With a faster 3MB on-die L3 cache, 6 instructions/cycle and 6.4GB/s of bandwidth, it is poised to perform at 1.5-2x of the current Itanium processor. There is an overview of how the Intel Itanium 2 at 1Ghz clock frequency will perform against the current Itanium 800Mhz and Sun's Ultra Sparc III RISC processor."

Comments

Re:AMD

[robburt]

At least the Itanium is large enough for me to put my coffee cup on to keep it warm. =)

No benchmarks

[KingKire64]

This is just a Marketing Piece put out by intel. All the "Benchmarks" are proposed Estimates. And why would a dinky website get a hold of something this "Big"? Dont know just questions.


Mod Me down Please

Re:No benchmarks

[Merlin42]

Personally I like how each page in the ppt presentations used a slightly different USIII ranging from 800 to 1050Mhz. Hmm smells like marketing picked out the best from a bunch of (simulated?) benchmarks. Everything was labeled 'simulated' or 'estimated'.

Nothing to see here folks please move along.

Well other than the slow death of competing high end architectures.

Lets see here we have:
SPARC ... still competitive I think
Itanium ... taking over
Alpha ... going the way of the dodo (im really sad about this)
PA-RISC ... transitioning to Itanium
MIPS ... never really liked them for big compute stuff, lets hope SGI can turn things around.
Power4 ... still competitive in performance but AFAIK to get a high end system you need to give your first born to IBM. And, IMO not really designed for the HPC kind of stuff I'm interested in.
Cray ... ? I've heard of some really cool stuff being developed, i'll believe it when I see it.

What eles is out there? I haven't really been in the market for a high end system in a while, but it feels like the market is shrinking and soon Itanium will be "the choice," unless legacy support is a concern .

Re:No benchmarks

[T-Punkt]

Even the graph has been done by a marketing guy.
They sorted the benchmark results in ascending order and the connected the data points of completely different and independet benchmarks by a line!

What shall the line tell you? The faker the benchmark the better the results? Or
"This is a line graph that doesn't make sense at all. But look: It shows an increase, increase is good, so Itanion 2 is good!"

Re:No benchmarks

[pmz]

Absolutely. From one of the slides "All projections based on Intel estimates[emphasis mine]...using...workload testing at Intel[emphasis mine]."

And, absolutely none of the benchmarks are substantiated with real data!!!

Only a fool would accept any of this presentation as fact. An even bigger fool would use this presentation in a decision whether to buy Sun or Intel.

Re:No benchmarks

[Anonymous Coward]

The reason they compare to different Sparc is because not all benchmarks are available for the newest one (Sun is playing the game too and is only publishing benchmarks where it serves marketing).

Note that Sun has been cheating on Spec, that's the only way they can make recent Sparcs look competitive. It's a shame that it will force every other vendor to cheat in the same way.

Re:No benchmarks

[binaryDigit]

Power4 ... still competitive in performance but AFAIK to get a high end system you need to give your first born to IBM. And, IMO not really designed for the HPC kind of stuff I'm interested in

Hmm, what makes the Power4 not suitable for your HPC needs. IBM seems to think it's more than fine since it's used in ASCI White (and anything else they sell for that purpose).

Re:No benchmarks

[joib]

Power4 ... still competitive in performance but AFAIK to get a high end system you need to give your first born to IBM. And, IMO not really designed for the HPC kind of stuff I'm interested in.

Huh??? What are you smoking? :) The HPC market is one of the primary markets for the power4, the other being big enterprise systems. In fact, CSC, the Finnish national supercomputer center is currently installing their new toy, a power4 machine which will have 512 processors when it's finished in september. FYI, that's 256 power4 chips, as one chip has 2 cpu cores. Anyway, the design consists of 16 fairly standard 32 cpu pSeries 690 refrigerator sized boxes. Currently I think they are connected with gigabit ethernet or something like that, but during the summer a proprietary IBM high speed interconnect will be installed. Total performance is estimated to be about 2.2 teraflops, more than 4 times faster than the old 540 cpu Cray T3E, and placing the computer among the fastest in Europe. Currently I think 6 nodes are operational...

BRAG MODE ON
And I have an account on that baby!!! *Drool* Wonder how many fps quake would get? ;-)
BRAG MODE OFF
No seriously, they naturally have a strict policy on what you are allowed to run on it. You have to fill out forms requesting cpu hours with project descriptions etc. etc. Anyway, my plan is to run ab initio calculations on it. Hopefully that is. They're having some serious problems, related to MPI, I think... Which has led to the fact that everyone is submitting the big jobs they planned to run on it to the old T3E, which is rapidly getting overloaded...:(

Re:No benchmarks

[joib]

At least some older T3s used MIPS processors (probably during the time they were owned by silicon graphics). Nowadays they seem to be concentrating on their new SV2 vector supercomputer, which has cray designed processors. I guess developing a successor to the T3E is not really viable any longer as clusters like the IBM system I descibe in the above post or even el cheapo Linux/Intel clusters do essentially the same thing as a T3E type computer at significantly lower cost. In fact, Cray sells Linux clusters these days (iron provided by Dell).

Re:No benchmarks

[javiercero]

NO! The T3s never used mips EVER! You are thinking of the origins 2K's that SGI relabeled as CRAY.

Re:No benchmarks

[joib]

Uh, yes. You're correct. T3:s are all alpha based. Guess it was just some idea I got when I was on a tour of the local supercomputer center a few years ago (they have a T3E with 375MHz CPU:s)...

Re:No benchmarks

[joib]

It is a great box but only if you use half of the CPU's. not enough cache to go around on the dual cpu boards so IBM has to disable half of them and sell them as 16 ways. Kinda sad.

And their switch architech for connecting them together is slow and has a relative high latency. But for what they are charging...

It is a SUN killer not a HPTC box. They have been telling their customers for the last two years that this box is for TPC not HPTC. HPTC low margin. TPC high margin. They want the HPTC customers to migrate towards linux.

Thier switch is not even really ready for prime time on these boxes.

Here you can see some benchmarks comparing the 32-way p690 with the 16-way HPC (ie. same machine but with only one cpu per power4 chip). The HPC achieves about 80% of the p690 performance. While this is impressive and suggests that the 2-cpu power4 suffers from L2 cache bandwidth I still think this is a semi-low-end configuration. When you want many cpu:s the internode bandwidth and latency is probably going to be more important. I.e. if you want say 512 cpu:s then 16 32-way p690:s are probably going to be faster than 32 16-way HPC:s, depending on your application of course. Not to mention reduced space, electricity and cooling requirements.

And also IIRC their switch so far supports only 16 nodes, although IBM claims it will support 128 nodes by the end of the year. And I wouldn't call it slow either, speed is 1GB/s (that's gigaBYTES, not gigaBITS), which is about the same as the T3E IIRC.

Finally, I'd say it's very much a HPC box in addition to a SUN killer. Take a look at the top500 list. There's lots of IBM SP:s there at the top isn't it? The p690 is essentially a SP with the 375 MHz power3:s replaced by 1+ GHz power4:s. Same goes for the switch. It's an upgraded version of the SP switch.

Of course all these IBM machines are seriously smoked by this new Japanese earth simulator machine, which is supposed to be 5 times faster than the current top500 #1, the ASCI white (an IBM SP with 8192 cpu:s).

On schedule???

[JFMulder]

...Itanium 2 processor codenamed McKinley that is on schedule...
Yea, that's if you forget the part about the first itanium being about a billions years late. :-p

Honestly, I am anxious to see what will come out of this war between AMD and Intel for the desktop market. Too bad they didn't have a comparison between McKinley and AMD's SledgeHammer, since they are destined to the same market.

And I would have posted earlier, but I was slowed down by the slashdot effect!!! :-)))

Yamhill?

[ultrabot]

But what's with all the stuff regarding MS urging Intel to use AMD's x86-64? Isn't the future of IA-64 rather bleak right now? Even HP apparently says that "market will decide" whether PA-RISC or IA-64 will be their future Unix platform... Which would not be the case if IA-64 was obviously superior.

Well, this can only mean good for Linux...

Re:Yamhill?

[kawaichan]

I think IA-64 should been seen as a long term investment for Intel, it's a good piece of technology (well, at least it was expensive) and it's just not up for prime yet.

x86-64 on the other hand, is a good solution for today's demand.

So who will have an upper hand 5 years from now? probably intel, but who knows, Ghz sells.

Re:Yamhill?

[Zeinfeld]

M$'s influence is negligible on all but the tiniest (1-4 way 32-bit) servers. So, on larger servers, you're talking about real OSes and huge applications for which 64-bit versions already exist (have for the best part of 10 years).

I don't know what you mean by a 'real OS' however having worked at the extreeme high end I can tell you that there are plenty of real fast machines with O/S that are in most respects (except performance) pure junk. And by high end I mean that some of the machines I worked on 10 years ago still outperform top end desktops.

The fact is that high performance machines are usually bought for fairly narrow purposes and as a result tend not to need a lot of an operating system. Back in the early 1990s an awfull lot of computationally intensive work was still being run on IBM mainframes running MVS and JCL which in many respects is not a whole lot better than MSDOS but you would never get the people who ran those piles of junk to admit it.

If you want high performance from a multiprocessor machine the architecture of the O/S does matter a lot, but may not determine the outcome. consider this analogy, when it comes to writing an optimizing compiler modern languages such as Eifel or Java give the compiler writer a heck of a lot more help than a language like Fortran. However until very recently many of the top benchmarks for compiled code were for Fortran compilers simply because brute effort could be used to compensate for poor architecture.

When it comes to the 'architectural features' required to make a multiprocessor machine work fast the cards are all with Microsoft. WNT was designed to work well on multiprocessor platforms and the design team were mainly DEC ex-VMS people who had a lot of experience in that area.

UNIX was originally architected for a UNIprocessor and there are a lot of design decisions that you just would not take if making it easy to run fast on multiprocessors was your objective. On the other hand this matters less than it might because there has been a lot of work since on compensating. The cost being that to make an SMP system work well under 'UNIX' often means having to use features that are proprietary.

The claim that UNIX has a better architecture than WNT is essentially as unprovable as the claim that vi is better than emacs. There are certainly still people in the computing world whose only experience of editing programs is with vi in line mode but will nevertheless post many gigabytes worth of posts to USEnet arguing the point. Having actually worked on O/S design, having used 20 odd O/S and having done system level programming on 4 (including UNIX) I can tell you that UNIX certainly did not get where it is on the strength of design merit alone. Many of the internals of UNIX are as confused and obfuscated as the syntax of the csh.

To port to a new 64-bit arch. with modern compilers and libs, it's not much more difficult than a recompile.

If you have nice 64 bit clean code that may be the case. The problem is that most people don't start from good code and even if they have tried to keep the code base clean they may not have succeeded. But remember that WNT has already been ported to a 64 bit architecture (Alpha) and although WNT is no longer supported on Alpha you have to believe that the compiler rules are still in place to detect code that is not 64 bit clean.

The issue that is probably more important for Microsoft is how .NET performs on itanium. With .NET they have the ability in theory to take applications and compile them for any architecture at installation time. I don't think there can be much doubt that this architecture is designed to support Itanium.

Re:Yamhill?

[Zeinfeld]

So, you've never heard of Solaris, then?

Slowlaris, oh yes, I had Sun sell me that when it was still a real heap of junk. Took the sysadmin a day each to configure four machines. He never did manage to get the multimedia (sound, video camera) features to work, nor had anyone else in the building. You might have heard of it, 545 Technology Square Cambs, Mass.

If they can't get an O/S to function two years after release it does not qualify in my book as superb...

Solaris kicks its ass clean out of the arena. Solaris has an extremely sophisticated internal threading model and scales nearly perfectly linearly to over 100 processors.

As you will note in my original post I differentiated the architectural contributions of Ken Thomson and co from later developers who just had the misfortune to start from the constraints of their design. It is possible to get a UNIX shell to run on a machine whose internal architecture bears little resemblance to UNIX.

The question is what the application developer is required to do to take advantage of the result. Does their application work in scalable fashion across all 100 processors if written for 'UNIX'? Or is it the case (of course it is) that to scale the application must either be 100 independent applications running on each processor or if it is genuinely one application does it have to be rewritten for the specific UNIX variant?

In the old days some processors would have a mode switch. So the VAX could emulate a PDP-11 but while it was doing so it wasn't doing the types of thing you would expect of a VAX, like virtual memory or such.

I have written code for machines with several thousand processors. Absent some development in hardware I may be unaware of nobody has a machine that can load more than about 16 processors that does not have some pretty significant limitations. There is a good reason for this, any given generation of silicon can support a certain processor technology and a certain interconnection technology. Unless you go to some form aof non shared memory system performance tends to top out at 8 processors. Yes I can believe that Sun showed off a machine that can calculate the mandelbrot set on 100 processors with perfect linearity, I doubt that it responds in that manner to all applications. Even exotic MIMD machines don't do that when they are purpose built for the application.

What iron has NT ever run on with more that 16/32 processors?

If Intel are right and they can provide 10 times the performance per processor over Sun then Windows need only run on 16 processors to kick Sun's ass.

However since the benchmark in question compared a 2 processor Intel machine with an 8 processor Sun machine the nature of the game should be obvious. Intel clearly chose a problem for which Sun's platform does not respond with near perfect linearity and the benchmark is obviously skewed in that respect. Even so, if Intel have the higher raw processor performance they have the engineers to build machines with more processors. Intel has at various times developed extreemly fast research prototypes (Bill Dally's hypercube being an example. Intel spends 3.2 Bn on processor research a year, Sun spends only 2 Bn on all its research.

Asking the number of processors NT runs on when Intel is only starting its 64 bit processor line is kind of beside the point. The real issue is where we will be in 3 years time...

Re:Yamhill?

[Zeinfeld]

"You might have heard of it, 545 Technology Square Cambs, Mass." Nope. Should I?

MIT LCS and AI Labs, if they can't get Solaris to work properly without spending a week doing it the problem is the o/s vendor.

We thought this somewhat strange since at the time SunOS had a realy good reputation on campus. Then we discovered that what people were using was a custom distribution cut by MIT...

As for hypercubes, the ones I saw were Transputer machines. Never wrote any code for them though. They only let me loose on the VAX cluster with 10 lines of FORTRAN *puke*.

You can't build a hypercube from a device with 4 links, its a geometric fact that you need 2 links per dimension. I did use Transputers, including some of the 1000+ node machines. Perhaps if you had some experience of the machines you are ignorantly blathering about your posts would not get modded down.

Your entire post confirms the argument I was originally making, that the term 'UNIX' no longer describes the system internals, it merely describes the application interface. However the application interface is still a major constraint on application design.

If you study the internal architecture of NT you will find that, despite M$'s continual assertions that UNIX is dead, outmoded etc. etc., they have come up with a kernel that, with every new release, is looking more and more like some of the big commercial UNIX kernels of 5-10 years ago.

Clearly from the tone of your post you don't like Windows, I have great difficulty therefore in believing that you have spent such a great deal of time examining the WNT internals that you can come up with such an assessment.

It would not be so surprising if WNT shared some of the concepts in Mach given that Rashid has been working for Microsoft for some time. However the architectural principles have from start to finish been based on VMS.

Re:Yamhill?

[Zeinfeld]

I spent absolutely zero time studying the NT internals, however my lectured did and he gave us quite a comprehensive description of NT, VMS, Solaris, SysV R3 SysV R4 and some of the really old stuff.

So you think your second hand account of some unspecified lecturer's opinions on O/S internals gives you the authority to pontificate on the issue?

I have first hand experience of VMS, NT and Unix internals. I suspect that either the lecturer was as opinionated as yourself, or more likely you didn't spend very much time listening to him either.

As for being a troll, if as you claim you are at Karma 50 and post at zero it sounds to me if a member of the slashcrew has come to an uncomplimentary opinion of you.

Re:Yamhill?

[Zeinfeld]

Well, from what he said, NT was nearly there, but not quite. It does have an advanced threading model.

NT does not have the UNIX thread model, or rather it has pthreads but the system apps all use a different model. This is not suprising if you know about thread models. pthreads is not generally considered the acme of good design. It is based on Tony Hoare's monitor design which is almost thirty years old now and Hoare introduced a much better design only a few years after monitors. However MULTICS was designed in the intervening years and when SMP Unix systems started to emerge in the late 80s the work was largely driven by the MULTICS work.

Opinionated is my middle name. However, I have to take his word for it. The guy is very experience in such matters and has written a lot of kernel-level networking code.

I don't know who your lecturer was but I'll bet he wasn't Tony Hoare, my Oxford Tutor who along with Per Brich Hansen and Dijkstra invented most of the concepts on which all modern O/S concurrency models are based. If Microsoft thought there was a problem with their threading model I guess they would call him up since he now works for them.

X86-64

[OS24Ever]

x86-64 may be more of a desktop migration point, but there are still plenty of IA64 type applications waiting in the wings from Microsoft, IBM and others.

There is always Linux64.

Not to mention the fact that many a beowolf supercomputer would like to be designed on a Itanium 2. There is one at NCSA from IBM with 800 some IA64 chips. They're just waiting for the Itanium2.

Re:X86-64

[roca]

The reason Itanic didn't ship in quantity is because customers don't want to buy a chip that's slow, hot, costly, and incompatible. Imagine that.

Yikes!

[delta407]

With a faster 3MB on-die L3 cache, 6 instructions/cycle and 6.4GB/s of bandwidth

Not to mention 130 watts of power consumption. And you thought Athlons were hard to cool!

Re:Yikes!

[Octorian]

According to Sun's website, the UltraSparc-III 900MHz Cu chip only has a dissipation of 65 watts.

run benchmarks in cache == FAST

[johnjones]

well I'm sorry but all the benchmarks seem to be cache hitters and so run pretty damn fast

real systems are about BANDWIDTH

memory bandwidth/latency is the reason AMD killed the P4 in benchmarks

lets see INTEL go up aganst a SUN on a large oracle DB then I will take notice

really this is where SUN make their money

regards

john jones

Re:run benchmarks in cache == FAST

[sql*kitten]

lets see INTEL go up aganst a SUN on a large oracle DB then I will take notice

Actually, Intel systems do pretty well, indeed, better than Sun running Oracle with a 3000G test database. And they do a good job on transaction throughput too.

Re:run benchmarks in cache == FAST

[VAXman]

memory bandwidth/latency is the reason AMD killed the P4 in benchmarks

There's definitely some confusion here. P4 (with Rambus) has much better bandwith than Athlon. On any memory bandwidth benchmark (e.g. Sisfot Sandra) the slowest P4 is faster than the fastest Athlon. The 533 MHz bus version of P4 does 4.26 GB/s, while the Athlon only does up to 2.4 GB/s or 2.7 GB/s. Latency-wise they're pretty much equivalent. They're the same using DDR, and 400 MHz Rambus is slower than DDR, and 533 MHz Rambus is faster than Rambus. Of course, for most applications, latency is more important than bandwidth which is why they're closer in overall performance than in memory bandwidth performance.

Re:run benchmarks in cache == FAST

[fliplap]

We should also remember that companies buy Sun for reasons beyond speed. Sun has legendary support, when something goes wrong, or when they make a mistake, it gets fixed == FAST.

Companies buy Sun because if for some awful reason a processor in an E10k dies they don't have to shut the machine down. The machine can be opened, still running, and the processors hot swapped.

The fact is, Sun machines are really really dependable, and thats what companies pay so much more for.

Re:run benchmarks in cache == FAST

[roca]

Yeah, instead of extending a crappy ISA, with IA64 Intel managed to design a crappy ISA from scratch :-).

Actually Hammer's long mode is a significant cleanup of the x86 instruction set. A number of rarely-used instructions and architectural features (e.g., segments) are removed, new general purpose registers added, various other features regularized (e.g., you can address the low 8 bits of every single register), and you can even ignore the x87 nightmare and use SSE2 as a clean floating point architecture ... it's faster than x87 too. (GCC can target SSE2 instead of x87 now!)

I just like to say "AMD's Hammer will crush Itanium" :-).

Yet more Intel vapor-hardware

[NetRanger]

I'm glad Intel is really optimistic about their processor. Now they need to either deliver the goods ... or just add more L1 cache and two more direct memory functions and voila , Pentium 5.

It seems that in comparison to finding ways to rev up the clock speed, PC-based innovation in processors has stagnated -- at least as far as those innovations that actually reach the market.

Perhaps I'm just picky.

Re:Yet more Intel vapor-hardware

[roca]

> The market will buy into it, indeed it is now.

Intel sold two or three thousand Itaniums, total. Pretty small "market".

Cool, but...

[A_Non_Moose]

does it make the intraweb go faster?

There's another variation on this story.....

[8127972]

At here.

Compilers

[Ted Maul]

I gather that the Itanium philosophy is to transfer the complexity to the compiler. The question is, how good are the compilers now? At the moment, it looks like a real bastard of a job putting together a decent one for any Itanium series. That VLIW stuff looks like it needs to be spot on every time to get the performance (don't do 3 fp ops in a row).

When I can run my C++ through an Itanium compiler and have it come out good, then I'll believe it. Benchmarks? Right.

Re:Compilers

[Junks Jerzey]

I gather that the Itanium philosophy is to transfer the complexity to the compiler. The question is, how good are the compilers now? At the moment, it looks like a real bastard of a job putting together a decent one for any Itanium series. That VLIW stuff looks like it needs to be spot on every time to get the performance (don't do 3 fp ops in a row).

Solid information content: 0
Repetition of things heard elsewhere: 10

Re:Compilers

[BlueGecko]

Actually, current compiler technology can probably optimize quite well for a VLIW architecture. The only catch is that for it to work properly, you will need to profile your code. But with that data, a good compiler that can figure out where to optimize shouldn't be that hard to write.

What I really don't get, though, is why no one is focusing on using JITs with these. It strikes me that this is the ideal platform for a JIT, where it can recode parts of the program on-the-fly based on where the bottlenecks are and so forth. I mean, wasn't this the whole point of a just-in-time compiler in the first place? IBM's Java runtime can rival C++ in speed if it is allowed to run for a reasonable length of time, allowing the code to become truly optimized, and since Intel is targeting this thing in a server environment where applications will run for a similarly long length of time I fail to see why they aren't going that route. This has the additional benefit that as our understanding of how to optimize for VLIW improves, the programs do not need to be recompiled, but instead can immediately get the benefit. (I am fully aware that Itanium is supposed to do some of this type of optimization itself, but current specs are utter crap, to put it lightly.)

Interestingly, Sun's MAJC architecture does exactly that, expecting that a JVM or similar virtual machine will run on top. I have no clue what happened to that chip, but it struck me that it had much better potential to kick ass than Intel's Itanium despite having similar designs precisely because it was designed for a JIT to be on top.

Re:Compilers

[roca]

It's hard to generate decent code for the IA64 so building a good JIT for it requires a very large investment. Furthermore the JIT compiler would probably be quite slow so it would have to run longer or achieve larger speedups for it to pay off.

Although a JIT would be able to discover and exploit behavior patterns that didn't show up until runtime (and therefore not exploitable by a static compiler), it's not a panacea. Lots of programs are unpredictable even down to the level of individual loop iterations. Such programs really need small branch penalties and hardware support for instruction reordering ... which IA64 doesn't have.

better parallelism

[Anonymous Coward]

Site is slightly slashdotted, and most of the data is in gifs. Here's a fact or 2:
Intel's claimed specint2000 and specfp2000 are both about 1.75x the 800MHz itanium. And this with only a 25% clock speedup to 1GHz.

They claim specint2000 is 1.3x Sun Ultrasparc3 1050MHz, and specfp is 2x.

Unfortunately, there is no indication of what the frequency headroom/scalability might be. The main point of the pentium4 architecture is to scale to 4+GHz. Can we assume anything similar for the itanium?

Re:What about I/O?

[Anonymous Coward]

Itanium2 still uses old and slow shared bus.

In 4-way configuration, each CPU gets only 1.6GB/s shared I/O and memory bandwidth.
UltraSparc III has 2.4GB/s memory bandwidth + some I/O bandwidth for each CPU.
Sledgehammer will have 5.4GB/s memory bandwidth for each CPU

SPECint????

[maitas]

I like the part where they said that Itanium 2 has 2x the SPECint performance of the original Itanium, since they never published it!! The SPECint performance for Itanium was so bad that only published SPECfp data!
It's just the same thing that happened when IBM published the SPECint/fp for POWER 4 processors. They only publish the data using 1 processor on the p690, so they run the hole SPEC benchmark suite un the 128MB SRAM cache memory, avoiding using regular DRAM. The easy way to see this is that they never published any SPECrate number, so to avoid showing that they don't scale as all processors start competing for the cache.
Sun USIII 1050MHz is almost 54% faster that USII 750MHz, as anyone can check going to the SPEC page (Sun Blade 1000 Model 1750 against Sun Blade 2050), with a 40% clock speed-up (this 14% increment is due to the compiler). This is exaclty the same processor at a faster clock, while Itanium 2 has more cache and a different architecture that Itanium, so a 1.5x to 2x speedup is less than spectacular, I will said.
For transaction processing, thay don't give any clue to show where they get the info from. While they expect to get the best OLTP number for 4-way systems, I don't think they will be able to surpass the AlphaServer ES45 MoDel 68/1000, which is by far the best 4-way system ever. What's worst, WLIW is know for been a poor performer for OLTP, and a great performer for floating-pont (that's why the only publisehd SPECfp!!). They never published any OLTP benchmark for Itanium (nor SAP, Peoplesoft, ORACLE, or even the raped PTC-C), so you can have an idea of how poor it is...
As of today, Fijutso PrimePower with 128 SPARC processors is the faster OLTP server ever (both SAP and TPC-C numbers!), with IBM p690 a close second for TPC-C and Sun SF15K a close second for SAP SD2-tier. Intel never showed in this kind of performance numbers, and Itanium certainly won't (unless while they keep running Windows).

Itanium 2?

[sharkey]

The first Itanic sank ALREADY?

Bad codename for a next-gen processor

[Ilan Volow]

Previous McKinleys haven't fared very well.

You are right.

[OS24Ever]

You are right, for windows 2000.

However, under Red Hat Linux 64 there are tons.

And note the release date of Itanium 2. Right along side Windows XP. There are supposedly 64 bit versions of SQL waiting to release with .Net Server 64.

Though honestly most people that ask me about 64 bit computing are Unix (Solaris, AIX, others) wanting to migrate to a less expensive hardware plaform running Linux to replace some lower end Sparc or Power3 boxes.

Though working for IBM I tend to work hard at the Solaris conversion than the AIX ones ;)

Re:You are right.

[OS24Ever]

From what we have been told at this point SQL is the main focus to compete with those high dollar installs of Oracle. I've not heard about Exchange but it can't be very far behind.

But SQL is the main focus to start the new major cash cow to compete with Oracle.

Keep in mind, I work for a hardware vendor though. They don't tell us everything.

Re:What about I/O?

[DarkEdgeX]

Uh, with a 4+ MB cache, the I/O bus and memory architecture aren't nearly as big of a hit on performance as they are on our piddly consumer processors with a mere 512KB of L2 cache. (And up until recently, 256KB was the norm!)

Properly written applications that take advantage of the cache (think video encoders that apply multiple filters on content already in the cache, for one example) are going to scream on this architecture.

Pfffft... One Gigaherz!

[Dr. Spork]

My chip runs faster than one gigaherz!

No, seriously, it seems that initial releases of the Hammer will have 2X the clock frequency of the McKinleys. I hope Intel includes an "Opteron rating" into the names of the various models just to help us keep things straight!

DDR, no Rambus!

[Dr. Spork]

From the slides, it looks like they intend to use this exclusively with DDR-200, at least around the launch. I think this is a wise move by Intel, and bad news for Rambus!

Logo....

[_ph1ux_]

I just think its funny that we have the official Intel (r) company logo for stories relating to Intel.... but we use the Borg Gates logo for anything relating to M$. Funny.

can you say ecache data parity error?

[Chaostrophy]

Sure Sun fixes problems fast. This one only got fixed if you had pull and screamed loud.

Sun did change the error message in later versions of Solaris 8.

I know a large company (you would know the name) who's production R3 system was off line for more than a day, mid week, because of this (crashed, and that trigered the problem that kept them down, thanks for the sucky support, EMC!).

Re:Why Is This Chip Necessary?

[topham]

I'm beginning to think I should start trading stock for a living. Every time I buy stock it drops in value. I figure maybe I should try to Sell-Short and then buy some. (not enough to cover it, but enough to cause the price to drop...).

Or, maybe I should just invest heavily in Microsoft.

Re:It better be quicker

[Paul Komarek]

"So it'll actually run even faster than the benchmarks let on once out of beta."

Bullshit. You don't know this, and Intel's history with new processor development suggests otherwise. Maybe this means that they finally have the heat issue under control, and can *finally* reach the clockrates they want to advertise. Maybe it means clock distribution is reducing chip performance to the point that heat isn't an issue. Maybe it means that you work for Intel marketing, and think you are repeating something you heard an Intel engineer saying at a party.

-Paul Komarek

Re:McKinley is a fast SOB

[roca]

> For high-volume secure electronic transactions,
> McKinley rules.

For the price of a McKinley you could buy a Pentium 4 and a pile of someone's crypto ASICs, and blow the McKinley away.

Performance improvements are conservative...

[LinuxParanoid]

The main point of the pentium4 architecture is to scale to 4+GHz. Can we assume anything similar for the itanium?

I don't follow this too closely any more, but I would presume they'll get to 2+ GHz, maybe 3 GHz, but probably not 4 GHz. A 4x jump is a lot to ask for without some additional redesign, especially if you are talking 4-way SMP running at those rates.

Given that they're claiming a 2x boost in SPECint2000 and SPECfp2000 from Itanium, on the same .18 micron process, the successor chips (Madison/Deerfield) on .13 micron should get them another 2-3x. Those are due sometime in 2003 I think.

--LP

Re:A fast CPU , oh wahey , the adrenaline buzz ...

[be-fan]

Its like a bunch of car bores sitting in a bar bullshitting for hours about
webber carbs vs bosch jetronic.
>>>>>>>>>>>
Here's your profile:
You use AOL.
You're VCR blinks 12:00
All you're Windows ME software still has the default settings.
You don't do your own oil changes.
You call a handyman whenever something leaks or breaks or burns.
You've never read the instruction manual to anything you own.
In total, you're a complete marshmellow who doesn't like to get his hands dirty.
Some people like tweaking, whether on cars or computers. More power to them!