Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
wmbz
thanks very much for your kind attention. However you are two days behind the curve in this thread. ;)
http://www.investorshub.com/boards/read_msg.asp?message_id=1325322
Joe
100% of server market is 1M chips per quarter.
1M Systems. (> 2M chips)
Even 20% in next quarter or 2 is still extremely optimistic (without suppor of HP and Dell).
Every double digit figure would be a sensation within this timeframe.
K.
Joe
Maintain mix of K7 and K8, with overall unit shipments steady, around 6 to 7M per quarter until 90nm, when full conversion to K8 will take place
Yes. That is my assumption as well. 6-7 M will be good for returning to Cash-Flow neutral operations if ASPs will improve by trading up the procuct-mix. Not really for black ink. Conversion to 90nm (waferouts) should happen from Q3/04 to Q3/05.
Although I can see AMD to return to profitability then, the latter would be very limited in my book without additional capacities, because even in an optimistic perspective for 90nm-capacity in Dresden I do not see big bucks in eps - unless we will see spectacular yields and/or Banias-like dense caches in this node. Realistically, it would probably only allow AMD to muddle through until 65nm node somehow, but then financing a share of a fab will be everything but easy.
Iaw, I dont see AMD capitalizing enough from K8 without additional 90nm-capacities to be confident for the longer term horizon of the company.
In essence, from an investment standpoint, AMD might provide a perspective for less limited profitability by means of a medium-term manufactoring-strategy if it would be interested in double digit figures for the common anytime soon.
K.
Joe
Because PGI compiler generated 64 bit code
Ok. The applications run for SPEC remain the same, right? That would indicate Intel-compiled 32-bit code is (presently?, still?) faster than PGI-compiled 64-bit code for SPEC2Kint, SPEC2Kfp and SPEC2Kfp_rate-applications. Hmm, for some other applications we have seen this phenomenon already.
I wonder how we should understand this?
Is it because compiler-technology for AMD-64 is still juvenile?
Or should we expect that a portion of the code out there will never benefit from being recompiled to AMD-64-code performancewise?
If IBM is pushing AMD64 bit code, I think it is more of an endorsement than if they used x86 code.
Under the assumption that desision-makers for Server-systems take notice of SPEC-CPU-scores in an appropriate way (if at all), yes.
Thanks
K.
Andy
"While the manufacturing constraint will certainly be true on the desktop"
If you can, any comment on what is intended to adress this one?
Klaus
wbmw spec scores
not what I expected...
Not sure the comparison you posted allows your conclusion:
The scores are from different plattforms. (Rioworks/Windows Server 2003/Intel/Compaq Compilers and IBM/SuSE SLES/PGI/gcc33 Compilers). Too many different things involved here to figure out what determines Scores to what extent. Interesting enough, 32-bit scores appear to be higher than 64-bit scores, something we have seen in other benchmarks already..
And then, assuming PGI Compilers indeed lower SPEC-scores, why did IBM use it for Spec reports? (Yes I remember it was me who posted about lukewarm commitment.... and on playing games around Opterons).
I admit I seem to miss a lot to understand what this all means in a broader context. Maybe individual CPU-SPEC-scores taken from a server system are pretty meaningless in the broader context of systems, maybe this all is just as misleading as some other benchmarks?
Would be glad if somebody could shed some light on this issue here. (Wonder how much innocence i still have to loose...)
K.
yb
From your numbers, I cannot make out (and dont expect as well) any marketshare gains. Seasonal pattern, nothing else.
For a 3% ASP up, too early to tell. Depends on the necessity of another sellout in September or not. K.
chipguy - why the heck did it just go on
a mini spending spree with a couple of questionable acquisitions?
Dont worry - They know exactly what they do. And there is good reason to not communicate it now. K.
yb
I'm not sure if I summarize that right :
As long as second hand Sparc is cheap enough to be a choice, we should not write off Sparc??
May I assume you would consider to buy a new Sparc if it is cheap enough to rectify it in terms of TCO?
In this case I would agree in not writing off SPARC: SUN has already announced to be prepared for individual pricing negociations.
K.
From your choices: 100.
I would not be surprized it is well below 100.
I would be very surprized if it is above 100.
K.
paul
MDT enjoys reputation for quality-modules and excellent price/performance in Germany. K.
windsock
"Advertising allowances are based on a schedule that relates to units purchased, ie volume."
Huh? Who told you that?
"RP didn't really work and the the small grocery stores that were the norm in the old days have faded away."
Now, we are together again.
K.
Techman - I wonder how fast they could make a Opteron/A64 system?
To begin with, no need to make - they could buy Opteron-Systems from Sanmina instantly. And make own things with it later, better said let it make by their suppliers in Taiwan.
For Athlon64, DELLs people in Taiwan could make it pretty quick. It is not difficult. K.
golfburn
You are certainly right: All prices are equal. Some are more equal, though. (By means of refunds for advertizing, projects, all you can think of. Completely legal, Robinson-Pattman act cannot prevent that.)
K.
keith, sgolds
From your link: The eServer 325 is designed for the demanding environments of scientific and technical computing customers. Iaw IBM decided to position this product for a segment targeted by Itanium. Although that is good to built up reputation, it creates very low volume. Not exactly where AMD intended to go with Opteron.
Any low-end server configuration is not even announced.
A blade server "under consideration" is exactly what I mean by lukewarm.
I understand pretty well what IBM plays here: Let Intel pay for NOT offering Opterons in the segments where the money is by delivering Xeons for fractions of listprices; enabling IBM to attack the DELL-territory. Yeah, I would have done exactly the same - as long as no competitor makes any move towards Opterons.
As for realistic expectations, you are right.
Its just frustrating to see a clearly superior product held in single digits MSS - or more precise realizing the big boys play games around it. Even more precise: Their customers let them go with it.
Again, you are right about realistic expectations. It is not exactly the first time that it happens in this industry, so I should have known better.
K.
Keith
IBM began selling the standalone eServer 325 using the Opteron last week."
Really? If so, I would call the offer indeed as a lukewarm offer, far from being really committed to Opteron:
I do not see reports to SPEC.org or to TPG.org about any Opteron-based standalone server-product-Scores.
What I can see is a 1U-rack-unit currently priced some 20 percent over Intel-Servers, without taken into account intended pricecuts for latter. TPC reported for a cluster-system available in November.
The message I receive here is: For those customers who insist long and hard enough on Opteron-servers - yes we have something we can sell you.
We can only hope there is a heck of a lot insisting. And biting the bullet of a considerable premium; as long as there is no competition at all from tier-one there is not much room for negociations on prices. Not exactly the competitive environment Opteron would need to make a significant dent into the lowend of the server market soon.
K.
keith
Could you pls take care of Elmers complaint?
Thanks
Klaus
paul,
yes i did. But made no copies.
The impression of the logos to me was positive.
Three logos in same style, one for Athlon64 (with some red in it) one for Athlon64 FX (black) and one for Opteron (green).
K.
yb
Conversion price is 7,37.
Within the last couple of months at around six dollars some short positions had been covered, realizing a dollar or so for it. Maybe that is what you consider upside pressure - which is limited to current levels where short positions are built up again - unless someone eats the blocks or the guys decide to let the common go up to make some money on the long side now. K.
yb I have no clue what the conversion rules are
Pretty simple: Holders of the notes can convert anytime.
AMD can buy back the notes not before December 2005.
So, shorting the amount of shares you have options to convert is zero risk for holders for the next two years. K.
burn2learn -
Thanks for clarifying a lot.
Humble questions to start with: What do LE and FE stand for?
Could you elaborate for me what excursions are (if possible like you would explain it to an eight year old girl)
Then, assuming both Intel and AMD run aggressive and non agressive processes, would you agree?
Your estimate on yield loss of 5-10 die per metal layer; assuming this is based on 200mm wafers that would make 40-80 dies for AMDs process (eight metal layers, correct?)
or something like 1/4 lost dies per wafer for metal-layer deficiencies.
This pretty exactly fills the gap between yield-models based on defects and capacities/output together with your other comments on losses.
K.
drjohn
np
See previous post...
Average power is well below 100 Watt, but my rate is well above 6 cents. However, it is in the same ballpark. That is what my point was in the context I made this remark.
K.
You must have very expensive utilitiy rates
Yes, indeed, we pay a lot more than 6 cents in Germany. We were coming down from 25 Eurocents in the 90s for industrial use by the way. Now, after deregulation we are at something between 10 and 20 Eurocents plus 16% VAT.
I used 15 Eurocents plus VAT and 50 Watts for average power consumption of CPU. In four years at 200 days at eight hours this adds up to 56 Euros. Which is what an AthlonXP 2000+ is for sale.
K.
semicon
Thanks for the explanation.
Hmm. I did not know that metal layers are referred to as backend manufacturing already. In this terminology, what do you think is the relation of defects frontend:backend?
K.
elmer
Thanks for taking your time to explain.
What I meant by "Timna" approach is Memory Controller and grafics on CPU-die: Two Dothan successor cores with large caches plus memory controller plus grafics. The chipset for it would be Twin Castle (nomen es omen). I dont see a reason this could not be made working in SMP, the basis for Dothan (PIII) works perfectly in dual configurations.
What I really dont understand is that yield models "for final wafer sorts" do not take the number of metallic layers into consideration: I can hardly imagine defects on and between these layers would be completely irrelevant. Could you (or somebody else) could elaborate on this? (Is this implicitely in different defect densities for different numbers of layers?)
As for the capacities - from the list of Intel FABs it is sort of hard to extract waferstarts for CPUs. On the AMD-side that is very easy. Then, again something hard to imagine for me is that Intel would have 100% better CPU-yields than AMD today.
(exactly this thesis some analyst confronted Jerry to last year in May. At this time and for 130nm process in retrospective I think it was accurate.) But now, with an obviously healthy process in Dresden, the numbers of CPUs sold and inventories versus waferstarts and diesizes result in only 60 of yield-figures than your estimates, even taking technology migrations into account allowing only a 75% utilization for volume production.
Somebody could fill this gap between my ears?
K.
dan3 - As it stands, much of that capacity looks useless since chips large enough to make good use of all that bulk silicon capacity (as opposed to SOI capacity) would dissipate too much heat.
That is what AMD expects and I as well. And yes, heat dissipation is a concern in server and especially in HPC segment - but obviously still not important enough to shift market shares very quickly in the volume part of it (2-way servers).
For mobiles (since energy consumption is just the other side of the very same medal) Intel already has an appropriate solution.
In desktop space I only see one regional market (Japan) where heat dissipation matters already today. In other parts of the planet, I never heard of people buying Barton because it consumes some 20 percent less energy than Northwood. The problem is you cannot see, hear of feel it unless you open the case and touch the heatsink. Very few people do this.
But we already talk about battery runtimes for mobiles - for the reason people do realize it. And i am sure it is only a matter of time when we will have power-consumption benchmarks for Desktop-Systems as well. And PowerNow! in Desktops. And hard-undervolting communities.
At current price levels, the energy bill for the power the CPU consumes over the lifetime is in the same ballpark of the pricetag of the CPU (assuming an office-system running 800 days in four years for eight hours.
We had this paradigm change in the automobile sector: Before first oil shock (1973 iirc) people talked about horsepower of their cars enginges primarily. Today, its fuel-consumption.
In the meantime, Diesel-engines became popular.
The question is when somebody is bringing the topic up for desktop PCs. Maybe we need another energy shock.
In essence, as for your thesis above: Centrino-Concept is at least a very strong indication Santa Clara is very well aware of the market's desires. Plus has a the marketing-power to influence desires to a certain extent. Plus has a product already to make Low-power dual-core CPUs from and utilize their fabs with it. No, i dont expect much useless Intel-capacities - unless the world is perceiving AMDs products as a clearly better choice at the point of sale.
K.
elmer
Well, based on "Dual-Timna" products (2cores, big cache(s?), graphics and memory controller on die and the fact "world class yields" are much lower than in the past (because process changes accelerate) an output of 10 Million CPU per quarter and FAB is probably closer than your 40 Mio figure. 200mm facilities could be used for Chipsets or other products or Flash or for possible demand increases or market share gains. If the latter wont occur some of them would have to be closed or sold after the old legacy fabs.
About the yield thing. Could you correct me in the following?
Based on Defect density, Wafer size, edge exclusion and yield model the resulting figure is line yield for one layer, right?
Now, defects in the metallic via-layers and in their connections to the silicon layer have to be taken into consideration as well, right? Plus everything that can go wrong in the backend, right? I am asking that because for the past and present Intel capacities yields as you assume for your 475 GDPW would allow to manufacture a whole lot more CPUs than they sell at full utilization.
K.
elmer - What is Intel thinking?
What do they anticipate that AMD doesn't?
That is what i asked myself for longtime. I dont think it is an anticipation but a strategy to outproduce competition using its capital and expected 300mm cost advantages. Not sure how exactly they intend to do that. Doubling cache sizes every year is not sufficient to explain the capacities they are building up.
Dual Dothan (or Dothan-successor)- Cores across the board?
Server first in 2004, Desktop next in 2005 (P4 will be Celeron then). With integrated memory controllers, and integrated grafics as well, possibly. Something like Dual-Timna.
And yes, interesting times ahead. At the very end it will probably turn out not really as a competition of approaches, but only a competition of timing:
AMD going to X86 and SoI first and will be forced to multicores later, plans to use strained silicon as well.
Intel going to multicores and strained silicon first, plans to use SoI later. And will probably go to X86-64 as well.
K.
wbmw petz Opteron xx6 Spec
Why dont you consider effects from using PGI 5.0 Compilers and ACML - at least for Spec fp - for your predictions?
K.
Edit: Petz I see you have mentioned it already.
yb I assume that will be a very hard resistance to cross?
Do you assume short interest and convertible shares from the automn bond are in equilibrium coincidently?
K.
Re: The snowball starts rolling
Thursday, July 10, 2003
Posted by - TelephonyWorld @ 11:44 pm PST
CEPOINT ANNOUNCES 64-BIT FAULT-TOLERANT 1U AND 2U TELEPHONY SERVERS
Nashua, NH, July 10, 2003. Cepoint Networks, LLC announces its Fault-Tolerant 1U and 2U AMD Opteron processor based turnkey Rack mount Telephony and VoIP systems. The systems complements the company’s RAID storage, NAS, SAN, SCSI RAID servers, SATA RAID and solutions for voice/audio data storage widely used in telephony and telecommunications applications. These computers and chassis are crafted with excellent cooling systems, dual hot-swappable N +1 redundant power supplies, and hot-swap storage drives capacity up to 1TB for 1U version and 2.2TB for the 2U version. Their blazing performance and reliability makes them excellent for wide range of telephony and call center management applications where 24 x 7 mission-critical operations is important.
Cepoint Networks, LLC is a systems integration and consulting firm specializing in applied computing products and services for telephony and telecommunications industry, Government, utilities and industrial markets. Products and services include turnkey rack mount telephony systems for voice processing, video/audio conferencing, message logging and telecasting. Portable computers for network engineers and technicians. Raid storage, SAN, NAS and servers for audio and video data archives and VoIP, custom hardware and software integration.
Petz
I agree. But I cannot imagine more than a small fraction of tens of thousands X-86 applications will be ported to IA64 in less than five years from now.
From the original Itanium-schedule (Workstations end of last millenium) this transition possibly could have been pulled by Intel. From today's viewpoint, I think it could only be done if Intel would pay up massively for porting.
K.
Red Storm and Top500
There is an interesting tidbit in Sandias pdf: Red Storms Architecture scales up to 30.000 CPU. That would propel Red Storm to #1 rank - at least in todays TOP500-list.
You know what my gut says? This is what they really intend to build next year already. I just cannot imagine U.S.A would miss the opportunity to take back the halo of this planet's fastest Supercomputer from Japan - for just 50 Mio bucks or so additional budget to expand Red Storm to 30.000 CPU.
K.
yb So, we should see dozens and dozens of small 96-Opteron supercomputers all around the world. It will be pretty cheap and the fastest thing you can get out of one rack.
Yeah, guess IBM will built dozens for its own computing on demand business - to start with. For a sales-position, these racks should be unasailable for any 64-way Itanium-rack price/performancewise.
As for the Concept of Red Storm, although not my turf (what you are discussing in detail exceeds my limits of understanding by far). But as far as I understand everyone of the architecture guys agrees in the following statement: (Posted by Mike Haertel on comp.arch recently, format edited)
The real dominant consideration in processor performance is latency to main memory.
(ed: Elaboration in context, i kept this message because of the implication that integrated memory controller of K8 is focussing this bottleneck as well) Most of the die area of most microprocessors is devoted to trying to hide this problem, and the bag of tricks used by CISC and RISC designs is the same. So it should not come as a surprise to anyone that the performance of CISC and RISC designs is also about the same.
From this viewpoint the way Red Storm is designed is nothing less than consequently adressing the bottleneck of computing as current state of the art.
K.
Haddock
If i would have been in AMDs, Sandias or Crays position, at the time of announcement that is what I would have communicated and/or assured in the contract. For at least these good reasons: To prevent from any doubts it is feasible and to keep some arrows in the quiver for the long time this topic will be in discussion.
When it comes to realization, I dont think Sandia will be bothered getting faster and/or less energy-consuming CPUs, dont you think? And AMD might have a viable interest to deliver the best they can at that time for this project.
As for costs of CPU, irrelevant, but Sandia has a fixed budget which will be shared by the parties involved according to the contracts already inked.
In a nutshell, you bet Red Storm it is gonna be more performant as disclosed at the time of the announcement. For the time being, just far too early to guesstimate how much more performant. Expect to become educated about that in some slices within the next year.
K.
Haddock - Not sure that Red Storm will be #1 in the world
I am not sure either. However, assuming 250-models from the consideration of my previous posting it should be at least very close to Earth-simulator (currently #1)
http://www.top500.org/lists/2003/06/
K.
Haddock
Red Storm is due to delivery in August 2004 afaik - more than a year away. So I am not sure if it is not intended to use 90nm-Opterons (250?, maybe coming on LGA-sockets and probably supporting DDR-2) already.
K.
sgolds
kpf, the plan has been to launch the value A64 on 90nm next year, along with mobile. There have been rumors of an earlier mobile launch, but I don't believe those rumors.
Not rumours anymore, AMD confirmed launch of mobile CPU in September (What I understand is they dont have the mobile-cpu specced as I would like one but want to sell lower speedgrade-CPUs - make it a µpga and call it a mobile CPU is not a bad idea at all)
The only thing that makes sense to me is for AMD to launch a high-end dual DDR version of the A64 early, if they can. I believe this can be done with socket 754. It is possible that, even though it has enough pins, there is a layout requirement which would force some pins used from socket 940. If this were the case then AMD would release it on socket 940, and differentiate by the fact it has only on HT channel.
And ECC/nonECC memory support.
However, I expect we will see dual DDR on socket 754.
Maybe, maybe not. Completely irrelevant for me which socket will be used this year. (btw I expect LGA-sockets anyway for DDR-2 next year as Intel already announced it from AMD as well. Seems to be the superior solution: Improved signal quality, less capacitors, smaller and probably cheaper as it will make µpga-packages obsolete as well).
K.
sgolds
And 939 pins? What is up with that?!
The differenciation of Athlon64 to Opteron 1xx?
Well, could be you are right at the end. However, some tidbits of the article (e.g. 512KB L2, afaik the very first time coming up in the press for the upcoming Value-model) fit well into what I would consider fairly probable - at least more probable as 256bit world and dog copies from older AMD-Roadmaps for it.
It is at least not impossible to combine the figures given to Mike by somebody who believes to know something in a way that makes sense.
K.
I personally dont care too much about sockets and pins at all. Just want to see a superior product to be launched in this quarter. However they do that.