mas,

SpecIntRate shows dual 3.6 Xeon 2MB a better performer...

... a draw in SpecFPrate

Are they running 4 instances?

Joe

Elmer,

Devastating roadmaps for AMD when you get out to 2007. (If correct).

The road map is pretty devastating for AMD even in 2005, with Intel running their Tejax chip at 10 GHz.

Another way to look at this roadmap is that Intel has one way out of downward spiral - Merom. If that chip turns like Prescott, it will take Intel down with it. Down from being a monopoly with out of this world margins down to being one of many competing semiconductor companies, surviving with razor thin margins.

Joe

Keith,

you sure you don´t mean 150mm^2 for 1MB L2 (512KB x2)?

Yes, that's what I meant, 2x512KB (not 512KB total).

Joe

Keith,

AMD could produce an Athlon 64 X2 with 512KB L2 (2x256KB), which running @ 2.0-2.4Ghz would offer very competitive performance, yet still have a die size of around 120mm^2?

Probably a little more than that, around 150mm^2, which is the size of Newcastle, which already sold millions. Rate of sales of 2 million Newcastle sized chips could be reasonably achievable, without limiting volume.

Joe

Kate,

Now we know where the code name Barton comes from.

Joe

chipguy,

Yet most customers won't buy Opteron. Its market share
actually dropped in Q4 while Xeon 64's share soared.

That would be an indication that clients really wanted AMD64 based Xeons, and were willing to wait to get them. Kind of blows the theory that customers were not interested in x86-64 out of the water.

Joe

chipguy,

Everyone likes to talk up power reduction but few seem willing to actually pay more for it in the blades market.

It is possible to get power reduction without paying more, by going with Opteron. Opteron offers solid power savings over Xeon.

Joe

wbmw,

So Win64 can make up for a fundamental deficiency of Win32. It's a non-intuitive little perk, but Win32 should not have been designed so restricted in the first place.

Just lke 640K, 2GB and 4GB didn't look at all restrictive at the time of release of 32 bit Windows.

Just like DOS extenders, there is a Win32 extender in form of PAE/AWE, but it would be foolish to invest money into its support now, when when the real thing is here.

Joe

wbmw,

I'm going to be frank and admit that I don't know the exact mechanics, but what is different from Win64 that no longer requires Windows to reserve 2GB?

If you plug a video card with some 256MB of RAM, Windows needs to reserve this memory for the card (apparently, for some reason twice, using 512MB right off the bat). Even though this memory is physically located on the hardware, plugged in the computer, Window needs to reserve memory for this hardware out of the 4GB (which is all 32 bit Windows has). When Windows 32 was written, Microsoft determined that 2GB for OS and 2GB for apps was a good breakdown.

There can't be any hardware limitations that requires 2GB of reserve, since Linux does not do this.

Linux does this as well, but they may have a different breakdown between OS and apps.

I can see this being a Windows deficiency, but why would it be a Win32 deficiency and not a Win64 deficiency as well?


With Windows x64, all the drivers for the hardware needs to be 64 bit - short term disadvantage, but the advantage is that the memory space for the hardware and the OS itself can be mapped to anywhere withing 64 bit memory space, with no need to touch app's 4 GB. (To be precise, probably a tiny portion of the 4 GB needs to be reserved, but nowhere near 2 GB.)

You mean that you or I could save money, but I'm pretty sure that most end-users who rely on a tier-1 vendor for their systems hardware aren't going to buy their memory from an off-the-shelf second source. Why do you think HP, for example, can charge such margins if their customers aren't inclined to buy from them?

I don't know what the practice is. When I worked at a giant company, that only buys teir-1, and I requested more memory for my desktop, I got Kingston, I believe.

If someone is buying 100 of these blades, and can save $1,000 per blade, $100,000 is something to think about even for the wealthiest companies.

It does hurt if the option comes at the expense of desired density or specific low power usage.

Right now, going with Intel for blades is at expense of desired density and power usage, since the only high end options are Opteron and Prescott / Nocona, as Dothan is not available from Tier 1s (I don't know if there are any blades based on more power efficient, but less capable processors, such as Northwood or Tualatin available).

As far as Prescott / Nocona blades, people buying and selling those whould have their heads examined.

But when Yonah is released, Intel should be more competitive in big part of the blade market, where 64 bit is not necessary. If there are dual core Opterons within Yonah power consumption enveolpe, they will probably be running at low clock speed, performing worse than Yonah.

For the rest of the market (mid-range power consumption, high performance), Opteron will be be a better choice, until either Pressler/Dempsey arrives and is is much more power efficient. If not, until Conroe.

Yonah will need to support SMP, IMO, to be considered for blades.

Joe

wbmw,

This is not as beneficial as you think. Windows requires 2GB of its own whether or not you are running in 32-bit or 64-bit mode. So you will not be able to run your 32-bit apps with 4GB of memory, unless you have 6GB of memory installed. And how much does 6GB of memory from HP cost these days?

Not exactly. Windows reserves 2 GB of memory for itself, it does not use all of it. You don't really need much more than 100 MB for the OS memory footprint.

Under Windows x64, Windows reserves some memory for itself, but it is outside of the 4 GB of application space. So if you have 4 GB, OS uses 200 MB, you still have 3.8 GB left for the app.

And how much does 6GB of memory from HP cost these days?

Of course, if you want to save money, you don't buy from HP, but from Micron / Crucial or Kingston. Here is Crucial pricing:

 
                                   HP              Crucial 
 1GB REG PC3200 2x512 Memory  [Add   $479.00]       199.98 
 2GB REG PC3200 2x1GB Memory  [Add $1,009.00]       551.98 
 4GB REG PC3200 2x2GB Memory  [Add $2,419.00]     1,551.98 
 

Shuttle has an ATI chipset based SFF: http://techreport.com/reviews/2005q2/shuttle-st20g5/index.x?pg=1

Joe

wbmw,

...And the number of applications that take advantage of more than 4GB.

Actually, there is a middle category, probably more important in the short run (than 64 bit apps able to take advantage of 4GB of memory). It is regular 32 bit applications that will go from 2 GB support to 4 GB support.

Joe

wbmw,

Besides, even HP Opteron blades don't support more than 2GB of memory:

Actually, they support 16GB of memory based on the specs:
http://h18004.www1.hp.com/products/servers/proliant-bl/p-class/25p/specifications.html

But it all depends on the number of DIMMS and size of the memory chips.

Joe

wbmw,

I think this is corporate bullsh!t designed to help AMD look good in the absence of DDR2 today. I think we are still missing a true apples-apples comparison between DDR and DDR2, and people are still holding onto the myth that DDR performs better, based on early results between Pentium 4 925x and 875 chipset comparisons.

DDR1-400 does perform better than DDR2-400. When it comes to DDR2-533, it depends on the application, whether it is bandwidth limited or latency limited.

Whe DDR2 gets to 667 and 800, it will clearly perform better, and if AMD does not have a processor out supporting it, Intel will have an advantage.

I don't think it is a huge deal for AMD to support DDR2, but the cost benefit is not exactly the same for AMD as it is for Intel. Fragmentation of SKUs cuts into benefits for AMD much more than for Intel, which, IMO, is the reason for AMD to not be on the forefront of this transition. In short, tiny (if any) benefit of DDR2-533 reduced increased fragmentation.

During the DDR from SDR transition, the benefit over SDR PC-133 was a non-brainter. With on-die memory controller, there is a risk to these transitions, that seem to happen once every 5 years or so. It is very fortunate for AMD that the current transition has much smaller benefit / penalty for being in front. If the transition was from DDR1-400 to DDR2-800, AMD would be in some trouble during the transition.

Joe

sgolds,

are they 4-way blades?

I don't know about IBM's, but isn't HP's BL45 going to be 4-way blade?

Joe

Keith,

I like this line:
HP also offers AMD based desktop solutions, which will soon be enhanced with even more capabilities

It shows only the Athlon XP based model. Hopefully, the 5150 model will be there soon.

Joe

Buggi,

Week 41 - "old CPU's" ...

It doesn't seem to say engineering sample. I wonder what 866 is though. I thought it would be called 865.

Joe

UpNDown,

shame they sold that extra server in 2003, wasn't it?

Yup. The growth rate, going from 1 Itanium server sold in 2003 to 233 in 2004 would have been 23,300% growth.

Joe

mmoy,

I just reinstalled the Express edition from February, and it said Beta 2 during installation for number of things (such as Framework 2.0 Beta 2, MSDN 2.0 Neta 2, etc.)

Joe

Bobs,

I think you missed the part of the story about the incredible growth Fujitsu's Itanium shipments are experienceing. 11,650% growth. From 2 servers in 2003 to 233 in 2004. With this kind of growth rate, in a few years, the entire surface of earth can be covered with Itanium die.

Joe

wbmw,

This statement makes no sense.

It makes sense when you consider how Intel planned to get to 10GHz, through very long pipeline, side effect of which was that the processor was more often and for longer stalled. Hyperthreading is there to fill these gaps with something useful, which is how we got hyperthreading.

P4 architecture (where hyperthreading is implemented) could not have been planned to go to dual core (the only one of the "merchang" CPUs, since it apparently will never be dual core (looking at the leaked roadmaps), like K8, Yonah and Mentocito. It is basically SMP on a die (or 2 dies), not "dual core" as most people would expect.

The only reason why there is going to be a dual core P4 is because there will not be a 10 GHz P4 (probably not even 4 GHz on 90nm), and P4 core is the only one Intel has that is suitable for current server and higher end desktop.

The idea that Intel prepared the market for dual core through introduction of hyperthreading is just PR on part of Intel executives, with no substance behind it.

Joe

wbmw,

DDR2 is the next step in DRAM technology, and only Intel is investing in it right now.

I think the difference (vs. Rambus) is that DRAMurais are actively supporting it.

Programming for Hyperthreading, meanwhile, is an evolution towards dual core. If Intel hadn't invested in multithreading development tools and helped optimize certain applications, we'd be even more behind right now, and dual cores wouldn't make any sense.

Hyperthreading was evolution to 10GHz single core, not dual core. Dual core is extention of dual CPU SMP, which has been around for a while. Hyperthreading is at best a detour.

Joe

wbmw,

I agree, and if the figures of 20% market share in the MP market are correct, then AMD has already had quite a lot of success.

I am not sure what exactly that figure covers (only the US, only x86?)

On the other hand, if much of the success was due to 64-bit memory addressing, then Intel is finally on par with EM64T.

Taking the whole market, x84 (or x86-64) with dual core chps and 64 bit OSs will make further gains vs. RISC processors.

I don't know if dual core is any more disruptive than 64-bits, but it could have a similar effect of gaining more market share in this lucrative market.

Agreed.

If they have 20% today, then 30% shouldn't be impossible with dual core and Sun's and HP's help. Much more than that will require IBM on board, though.

IBM has conflicting interests between Power architecture IBM needs to protect, costs already sunk into Xeon chipset, and having competitive solution for its customers. Opteron may end up victim of internal IBM politics (although 4-way Opteron server has been mentioned several times).

Joe

chipguy,

ROFL, twice the wafer utilization and nearly twice the
manufacturing cost for a feature the mass market will
not pay a lot more for even if several speed grades were
not lost. Yep, sounds like a surefire recipe for huge AMD
profits all right. ;^)

I doubt dual core desktop will get much more than leftover capacity of Dresden fab, and will likely not have a material impact.

But dual core Opteron should result in a gain of market share in server market. 4 way, dual core Opteron server with 64 bit OS is very much a disruptive technology for that part of the market.

Joe

Keith,

I wonder if they have 2 mobos (754 and 939) or if AMD is offering s939 Semprons, or some weird Socket 754 A65 3500.

Just for comparison, the new notebook line (6000 something or another) has similar processors...

Joe

Klaus,

-->65nm takes a while longer than originally planned.

Not really that late. 2nd half of 2006 for volume is kind of what I expected, with initial volume at 90nm

--> Cell is fabbed at 90 instead of 65 (commitments would not allow to postpone it)
--> Cell eats up Fishkill capacities

I doubt Cell will eat up the entire capacity, and if anything, rather than canceling everything else, if Cell volume is huge, I think IBM would farm out some of it to AMD.

--> Lack of development capacities in East Fishkill
--> Fab36 has the right toolset and is a member of Consortium.

I think process development stays in East Fishkill. It depends of some key people, and you can't just move them to another continent without losing significant number of them.

Also, IBM is building another fab in the same complex.

Joe

But the Express editions are all Beta 1. Some later (February) some earlier (June 2004). Beta 2 has not been released AFAIK.

Joe

K. Tx. EOM. Joe

mmoy,

Where did you get Beta 2? MSDN?

Joe

UpNDown,

Pacifica powerpoint presentation is now up at:

Good thing i had my Venti size coffee, otherwise, the subject of virtualization would surely put me to sleep by now, and I would not be able to reply. I think AMD and Intel should think about some joint marketing venture with Starbucks, when they push their virtualization enabled CPUs

Joe

Tenchu,

Actually, I've seen a lot of them, including the very first ones when AMD announced x86-64. Integrated memory controller, HyperTransport, 64-bit extensions, but nothing except a one-line mention of dual-core.

Here is the link to the original presentation from 2001:
http://www.amd.com/us-en/assets/content_type/DownloadableAssets/MPF_Hammer_Presentation.PDF

Any technical guy would know that ANY processor design capable of multiprocessing can be turned into dual-core.

If you are talking about using the facility for multiprocessing, which is the FSB, yes, but that's the dumbest approach of all, something Intel is using in P4 based Smithfield. The other approaches, namely Yonah, K8 and Mentocito are smarter, where the arbitration happens inside the multicore CPU die.

K8, Yonah and Mentocito are true multicore chips. To call Smithfield a multicore is kind of stretching the definition. It is more like SMP where the 2 CPUs either sit on the same die or are close next to each other in a multi-chip package.

Of course, CSFB aren't technical guys, so they probably mistakened AMD's headstart in dual-core design as proof that K8 is more "well-suited for dual-core."

Well, both are true (head start and well suited).

AMD did have a head start, since the implementation of the plumbing for multicore has been in silicon for 2 years. Unfortunately, dual core was targetted for 90nm node, which has arrived at AMD somewhat later than at Intel.

As far as well suited, P4 is obviousley not at all suited for dual core (or Intel would have made a real P4 multicore).

As far as Yonah and Mentocito, I guess they are well suited, but if they were better suited for future multicore CPUs, I guess AMD would be trying to copy their architecture. Since it is the other way around, Intel's opinion is that Opteron architecture is the one that's best suitede. Which happens to be the same as the opinion of the CSFB guy.

Joe

Tenchu,

There's nothing in the K8 architecture that suggests it is more well-suited for dual-core than the P4 architecture.

Have you seen any of the K8 presentations?

If anything, it's the opposite. It's easier for Intel to go to dual-core just because of the shared FSB, which allows Intel to simply "glue" two cores together on the FSB pins. AMD can't do that; instead, they have to add logic to make their dual cores share the HyperTransport links.</i.

The logic has been there for 2 years.

In short, CSFB is spinning things negatively for Intel. I highly suspect ulterior motives here ...

Most people have ulterior motives, but it just so happens he probably has a good case on this one.

Joe

chipdesigner,

It's a DTR notebook (8 lbs), and Winchester desktop parts are low enough power for that. Interesting choice. Another HP-only thing?

The CPU parameters do sound like those of a Socket 939 CPU, don't they? But I have not found any reference to Socket 939 or dual channel memory. Dual channel memory should help with the integrated video.

Joe

mmoy,

It is not really low end, high end wideo. It is the same ATI integrated video. In one case, it uses only CPU-memory bandwidth, in the other case, it has additional 128MB of local memory (for increased bandwidth). Great option for $25, IMO.

Joe

I see this new zv6000 notebook on HP site:
http://www.shopping.hp.com/webapp/shopping/computer_series.do?series_name=zv6000_series&catLevel...

This is the one with ATI chipset. I am a little confused about the choices of CPUs. Sempron or Athlon 64, but the Athlon 64s appear to be Socket 939 desktop models. They also claime HT speeds of 1000.

Joe

Klaus,

Why should Sun challenge Andy's Galaxy platform with Horus or vice versa?

Why are you assuming Galaxy is not based on Horus? The only thing we know is that Serverworks will be providing chipset, and there have already been some rumors / demonstrations of Serverworks chipset, which kind of parallels AMD;s 8131 / 8111. There has been no info about anything from Serverworks that would halp connecting more than 4 processors, so it is possible that Horus will be used.

Joe

Klaus,

Sun would make an ideal customer for Horus, to replace their higher end Sparc lines, as would of course Dell, which does not have a strategy for a high end.

Joe

Klaus,

Longerterm, while I believe Horus chief architects move to AMD will probably lower the expectations for this architecture, Rich has what it takes to initiate and drive development of a suitable big iron K10 platform architecture for deployment in some two years or so.

I was wondering if it is overall a net plus or minus for Rich to move to AMD. I wonder if Newisys has any customers for Horus.

Joe

Tenchusatsu,

The numbers are never "small enough" because among other things, the vendors also have to shoulder the burden of the recall, and the blow to Intel's reputation is immeasurable.

Suppose 2 possible scenarios were weighed:
1. Intel's potential liability of $1 billion to Intel, while being competitive with AMD
2. Not being competitive with AMD and giving tham 0.5 to 1 billion worth of profits

Which one do you think Intel would pick (in a heartbeat)?

Joe

combjelly,

I think HP just needs to bring up their web ordering to the level of Dell, as far as selection, service, timeliness of deliveries, and that portion of their business will grow. I have ordered things from HP directly, mainly components, and it is not the most pleasant experience.

Joe