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.
Edgar -
Do you guys think the .38 earnings est is now blown away due to the Intel news or are we just going to beat by maybe a dime?
I am not a good predictor of AMD's earnings but I really don't see anything to change earlier forecasts.
Spokeshave -
You didn't direct this post to me but something tells me you won't mind if I jump in<G>.
Do you not think it is possible that Intel got caught up in the "irrational exuberance" of the tech bubble and made plans and commitments in anticipation of continued stellar growth that never materialized?
I don't think this is the case anymore. F24 was put on hold for just such reasons but building was restarted earlier this year.
Conversely, is it not possible that yields are not as good as they were two years ago? Is it not possible that there is a mix of the two that explains the aded capacity that seems unused?
Both of these are entirely possible, I've never argued that they aren't. I've only argued that you can't conclusively determine yield problems to be the cause of your discrepancies. I never comment one way or the other on Intel's yields.
Edgar
Anyone care to agree that this present overbought short squeeze has run it's course?
Are you shorting here?
SNO -
You did a nice job setting up the spreadsheet, so why not use meaningful data?
Here you go:
http://icknowledge.com/misc_technology/die_calculator.xls
The default defect density is .50/cm2 but most insiders would consider .25/cm2 to be world class. If you consider either Intel or AMD to have world class yields then consider using the lower number or different numbers for either company.
This is very good as a first approximation.
Sqrt_negitive_one
No Elemer, there[sic] not wrong. They are right based on the percentage yield I picked. You can pick out any percentage yield you would like and do the numbers in a spread sheet.
Sorry, I thought you were trying to be accurate. Silly me...
But for those who want to cost model real silicon, you will need to use different yield percentages for different die sizes.
Spokeshave -
That's a "Are you still beating your wife" question. You are assuming as fact that there is low Opteron output. I have nothing other than Inquirer statements to support that position. And, how do you quantify "low"?
I didn't mean it that way.
The fact that there are retail packages available in copious quantities tells us only one thing - AMD is able to make enough to satisfy both OEM demand and "Pricewatch" demand.
Exactly.
kpf -
What I remember from the recent update was that Intel was seeing unexpected demand and as a result they were seeing spot shortages across several product lines. They didn't identify the products and I don't see it as an indication of capacity constraint but simply the result of the "unexpected demand".
That's my recollection.
Sqrt_negitive_one -
I hope this helps in your quest. This is something I worked up in early 02. (sorry about the lost formating)
.
.
Dice count 208 264 340
Yield % 0.75 0.75 0.75
Sellable 156 198 255
No need to go any further. You numbers will be wrong from here because the yield percentages will go up as the die size goes down. They will not be the same.
8-/
So you dont believe in the statements of AMDs and Intels management about utilization: Both say CPU capacities are fully utilized. (AMD says FAB25 is underutilized, Intel says Chipset is underutilized.)
I follow both companies rather closely and I don't remember Intel saying they are at capacity for CPUs. Where did you get this?
DDB
The bulk process has fine yields
Please provide some data to support this.
VIA Technologies posts NT$1.1 billion Q2 loss
BLOOMBERG
Friday, Aug 29, 2003,Page 11
VIA Technologies Inc («Â²±¹q¤l), the world's second-largest maker of chipsets for personal computers, had a third straight quarterly loss after costs rose and competition with Intel Corp and other rivals eroded its market share.
Net loss was NT$1.1 billion (US$32 million) for the three months ended June 30, compared with net income of NT$46.4 million a year ago, the company said in a statement.
Sales, reported earlier, fell about a quarter to NT$4 billion from NT$5.3 billion.
VIA's costs increased after it settled a patent dispute with Intel, the world's biggest chipmaker, in April. The Taipei-based company's non-operating expenses climbed more than half to NT$1 billion in the second quarter from NT$615 million in the first.
"Non-operating expenses were higher than in the first quarter" because of licensing fees, said Cheng Ming-kai (¾G©ú³Í), an analyst with CLSA Asia-Pacific Markets in Taipei.
"I don't think this will be a one-time payment," Cheng said.
Intel settled 11 patent lawsuits pending in five countries with VIA on April 4.
The two companies, which didn't disclose financial terms, signed a 10-year agreement to license each other's products and Intel agreed not to assert its patents on some microprocessors for three years, the companies said then.
Intel gave VIA, which also sells microprocessors, a four-year license to design and sell chipsets compatible with the US chipmaker's processors in exchange for royalties.
VIA lost sales to Intel and Taiwan rivals after they introduced new products early this year, analysts said earlier.
Chipsets work with processors such as Intel's Pentium 4 to manage personal-computer functions such as memory and graphics display.
sgolds -
Elmer, you seem to miss the point. Where is AMD suppose to introduce new product into manufacturing? How are they suppose to convert any fab lines without taking down production during that conversion?
You're right but I'm not sure who that argues for. You guys are allocating to development whatever space is necessary to make the remaining fab capacity/output equate to good yields. If we are supposed to continuously write off a large portion of AMD's capacity for development then AMD is stuck in a position similar to trying to change cloths in a tight phonebooth. Development will always be ongoing. AMD will never have the full output of a modern fab, even with decent yields. AMD will never be able to capture the market share they seek and all is lost. I don't believe R&D consumes more than 5-10% of capacity and cleanroom space has been increased since AMD announced 5500WSPW capacity. 5500WSPW is a reasonable estimate for production capacity.
Spokeshave -
The real question is *why* do you disagree with this number. I already accounted for your objections and made the model even more conservative as a result. Do you still have objections that you did not state?
You have done the best (only) job of estimating utilization potential but the unknowns are just too great imo to arrive at any meaningful conclusion about yields. Yes, the capacity has increased but we have no idea how close your estimates are to real life because of all the diverse products and unknown volumes. When I estimate AMD's yields I make a point of stressing that we can only draw conclusions when demand is high and fabs are pushed to capacity. Is AMD's low Opteron output a yield problem or just low demand? We can't know for sure. Same goes with Intel. To date we've seen no indication of capacity constraint so perhaps the question should be why does Intel have so much extra capacity? That's a question I don't have an answer for.
Thanks for your thoughts. You're the only one who was willing to make a case.
Sopheshave -
You don't even know how to accept a compliment...
We all know that there has not been a unit demand increase of even 10% since 2001
I don't agree with this at all. I think Intel's overall unit demand increase could easily exceed this.
So, overall, there is still about 45% capacity unaccounted for.
I disagree with this number but why can't you just chock this up to excess capacity? Who said Intel is capacity constrained?
Spokeshave -
I didn't think you had it in you. Bravo! Someone actually took the time to make an logical estimate. That doesn't mean it's entirely correct but you actually took a structured approach! I'm almost getting choked up...
Let's look at your analysis:
Intel Fabs. Good list but there are still some missing and some adjustments to be made. You left out Fab8 & Fab23, both 200mm fabs so add some capacity. One of those very large Fabs is currently offline production wise because 90nm product has not resulted in revenue as of this date. Subtract some capacity. Another one has not reached full capacity yet but is still ramping (you mentioned this). Another one is shutting down for 300mm conversion so it's ramping down, subtract some more capacity.
According to In-Stat/MDR, Intel should approximately double its wafer area production, versus 1Q01 production by the end of 2003, due to the rollout of 300mm wafer capability.
I'd like to see this quote. Do you have a link other than the one at the bottom of your post? It is very non specific (You have to pay big $$$ to get the full report). I suspect it includes 2 or 3 300mm fabs that are not currently online. #1 D1C is 90nm and not to be included in today's volume. #2 F24 comes online in early 2004 and is probably included in the In-Stat/MDR numbers. #3 D1D is not a production fab. 3 300mm fabs would be eqvilalent to about 7 200mm fabs. That could easily skew your numbers.
Since 1Q01, average die size has decreased about 20%
What do you base this on? Got a link? I hope you're not going by DT processors alone. How do the chipset die sizes compare to Q101? Intel makes 100s of millions of these each year and you need to factor that in. Also Intel is now the largest graphics supplier with their larger integrated north bridges. They weren't in Q1'01. You also have to adjust for very weak demand in '01. I see you mentioned this as well.
Throw in Madison and Gallatin with their very large die and you might be mistaken in your assumptions on average die size. How has Intel's volume in the PDA market changed? Networking, wireless, automotive? Are Flash die the same size as before?
I give you credit for being the only one to make a logical attempt to estimate capacity but I think the unknowns are far to great to draw any meaningful conclusions.
8-/
We can never know their actual yields but we can tell what's coming out the backend and in the last couple of years it hasn't been what it could be if they were well running volume fabs
I know this is getting a little old but let it be pointed out that you don't know how many fabs Intel has, how big they are, where they are, what processes they run, what products they produce, what their die sizes are or what their volumes are, yet you think you can draw conclusions on their yields.
You've certainly shown you analytical skills.
sgolds -
With all the products AMD has going at Fab 30 at one time, it is impossible to know what proportion of the factory is dedicated to production at any one time, unless AMD tells us.
This is one of the highly contrived explanations, in my opinion. The other one was "just bad planning". If AMD is going to have a large percentage of their capacity offline for process development plus running multiple processes on the remaining then they're never going to hit their stride and produce the volumes needed to reach profitability. That's my opinion. You're right, we can never know their actual yields but we can tell what's coming out the backend and in the last couple of years it hasn't been what it could be if it was a well running volume fab.
SGI keeps rolling out the performance. I bought a couple thousand shares yesterday @$0.98. Can't lose too much.
SGI Altix 3000 Performance Lead Rolls on With Latest SPEC
PRNewswire
August 28, 2003 (9:30 a.m. ET)
SGI Altix 3000 Performance Lead Rolls on With Latest SPEC Benchmark Results
<more>
http://www.eetimes.com/pressreleases/prnewswire/95746
SilentBob -
At least one poster has claimed AMD's yields are extremely poor based on supposedly obvious "facts" with no hue and cry from a certain group of posters. Funny how Intel's yields are now incalculable using those same "facts". LOL
I don't remember anyone stating for a fact that AMD's yields are poor although I believe they are and I think I've made a pretty good case for it. In AMD's case they have a single Fab with known capacity. We know the die sizes and we can predict the potential output from AMD's Fab assuming industry standard yield. What we have seen in the recent past was that AMD never was able to sell more than about half their theoretical capacity, even in times of high demand. Several attempts have been made to explain this without involving poor yields but they are quite contrived and I am of opinion the poor fab output combined with several factors is much more easily explained by poor yields brought on not by high defect density but the need to push performance to the raged edge of their processes capability, just to compete.
As for Intel, nobody here seems to be able to list their fabs, their capacity in wafer starts per week, the processes they run, all the products they produce, their die sizes or the volumes of each. Yet you see the 2 as equivalent. I think there's a big difference and I'm sorry you can't see it.
gollem -
Elmer, right.
You said Intel has enough capacity to make every chip on the planet.
I asked for data, you had none.
I asked for an estimate, you had none.
I asked for a simple guess, you had none.
Thanks for nothing
Shouldn't I be saying this to you?
Gollem - I don't have all the information, between us we have quite a bit of information on this board. You think intel doesn't have too much capacity to be consistent with high yields, I do, neither of us is certain. If we try and estimate total capacity and total needed capacity that would be an interesting exercise.
You said "intel has enough fabs to make every chip on the planet" so I assumed you had data. Now you are saying you don't even have an estimate? How can you make a claim of fact when you can't even make an estimate?
How about doing this, just make some guesses? Guess how many fabs Intel has and guess their capacity. Then guess how many products Intel makes and their approximate volumes.
When you're done with this we can estimate the rest of the world's volumes and see if you were close.
gollem -
intel has enough fabs to make every chip on the planet, assuming they have the great yields you say they have.
I've seen this claim made before and when challenged no one ever presents the numbers to show how it can be true. Since you've made this claim I must assume you have the data so please share it with us, we've been waiting a long time to see it. Please list Intel's fabs, their capacity in WSPW on each process node, the products produced and their volumes, along with references supporting your numbers. If you include the yields of each that would be nice but we can easily derive that from your data. This will prove you really knew what you were talking about when you made that claim.
I'm hoping this post won't be deleted for being an abusive personal attack but I'll just have to take my chances...
Lab Soups Up Linux Supercomputer
By Robert McMillan
IDG News Service, San Francisco Bureau
(08/27/03)—A 2,000-processor Intel Corp. Itanium 2 supercomputer at the U.S. Department of Energy's Pacific Northwest National Labs (PNNL) has edged out Lawrence Livermore National Lab's Intel Xeon-based Multiprogrammatic Capability Cluster for the title of world's fastest Linux supercomputer, according to PNNL.
PNNL on Tuesday announced that it had completed an upgrade of the 1,400 1.0GHz Itanium 2 McKinley processors in its William W. Wiley Environmental Molecular Sciences Laboratory supercomputer in Richland, Washington, boosting the system's peak performance from 6.2 trillion floating point operations per second (T FLOPS) to 11.8T FLOPS. The new processors run at 1.5GHz and are based on Intel Corp.'s follow-up to its McKinley design, which is called Madison.
"It's about 11,800 times faster than the average personal computer," said PNNL Molecular Science Computing Facility's manager of computer operations, Scott Studham. "Most computers have between 250M Bytes and 1G Byte of memory. This one has 7,000G Bytes of memory."
Linux has emerged in the last few years as an increasingly popular operating system for the highly technical supercomputer market. In the last month, Dell Inc. announced plans to build a 17.7T FLOPS Xeon system for the National Center for Supercomputing Applications, and IBM Corp., Fujitsu Ltd., and Cray Inc. all are building Linux supercomputers in the 11T FLOPS to 40T FLOPS range.
PNNL's upgrade process took just over a month, with a team of 10 Hewlett-Packard Co. employees on site unpacking and installing about 250 Madison microprocessors into the Labs' McKinley-based rx2600 machines each week. "On a weekly basis, a semi truck with processors would show up," said Studham, who claims to have developed more than a passing familiarity with the CPU upgrade process. "I can personally tell you that there are four screws required to take out an Itanium 2 CPU," he said.
The 3,000-square-foot, US$24.5 million system will be used for a variety of computationally intensive tasks at the labs, such as studying basic chemistry and biology, and modeling how leaked radioactive material might move underground.
For this kind of science, the Itanium 2's floating point performance of 6 billion operations per second made it a better fit than AMD's rival Opteron processor, Studham said. "It was important for us to build out of the fastest processor we could get," he said. He estimated the labs would have needed 1,000 more processors to achieve the same level of floating point performance with an Opteron-based supercomputer.
Thread - Thought I'd show you the response I got from Matt when I told him his censorship policy is unreasonable -
Unreasonable? You constantly are attacking people. It's simply not necessary in a conversation.
I don't think I've ever attacked anyone here but apparently Matt thinks I'm "constantly attacking people". For those who think I've attacked them, my apologies. For those who wish to be allowed to interact as adults, you might consider other options.
EP
Thread - Thought I'd show you the response I got from Matt when I told him his censorship policy is unreasonable -
Unreasonable? You constantly are attacking people. It's simply not necessary in a conversation.
I don't think I've ever attacked anyone here but apparently Matt thinks I'm "constantly attacking people". For those who think I've attacked them, my apologies. For those who wish to be allowed to interact as adults, you might consider other options.
EP
Boruas -
I have not been suspended. I have simply decided that the level of censorship here is so ridiculous as to not allow adults to carry on meaningful discussions. The posts I was admonished for were not abusive, vulgar, argumentative or in any way objectionable, imo. That's what makes it so frustrating. The pettyness of the administrators simply leaves me no choice but to move elsewhere where adults will be free to discuss technology, investments and interact as adults without the smothering overbearing parental control practiced here. I am in complete agreement with deleting vulgar, abusive, off topic, threatening or otherwise offensive posts. The administrators here have gone far beyond those standards and imposed a kindergarden level of censorship that destroys free discussion.
Thread - I am being accused by Matt of making abusive posts. I feel this accusation is outrageous and totally without justification. I'm out of here for awhile. I don't need this.
wbmw - It's still too bad that they lost the one against Intergraph.
I believe portions of that suit are still ongoing.
Broadcom to pay out $60M in Intel lawsuit settlement
Companies enter into cross-licensing agreement
By Tom Krazit, IDG News Service August 08, 2003
Intel and Broadcom have agreed to put their legal differences aside, but the settlement of the lawsuit filed by Intel three years ago will cost Broadcom $60 million this year, the companies announced Friday.
YB - Matching Xeon price is the easyest way to get almost optimal performance. I think they are right, at least it's good to try.
This is very risky. Should AMD be forced to lower the price, relative to Xeon, they're going to look really bad.
NaS - and now that the two of you agree that 10% + 20% = 30% can you stop and let the little people know what AC stands for?
Well technically it stands for alternating current but what it means is that it isn't a static or parametric test, it's functional and hopefully at speed. If we really want to be kosher it is a full speed functional test with timings set to datasheet specifications. Setup and hold times are as specified, output delays are tested as specified, Vcc is as specified and pins are loaded as specified. Additionally all parameters are guardbanded to account for the inaccuracies of the tester and even temperature is guardbanded to allow for environmental inaccuracies. We should even throw in guardbands to allow for hot electron degradation and similar P-Channel effects. An example would be that your test equipment has a guaranteed accuracy of 125ps, meaning that the signals can vary by +/- 125ps. To correctly test setup time you must tighten the specification by the total inaccuracy of the test environment. Not only is the signal in question inaccurate by 125ps but the latching mechanism, usually a clock, is also inaccurate by 125ps. You must tighten your specification by 250ps to guarantee accuracy. Make sense? Frequency will degrade too because of effects described above so that needs a guardband as well. This is the reason that all processors must overclock. They have to because they will degrade over time. Hope this helps.
Chipguy - 10% of device run at 2.2 GHz or above, 20% of devices run at between 2.0 and 2.199 GHz. The remaining 70% of functional devices run below 2 GHz. Therefore the AC yield testing at 2 GHz is 30%.
OK, I guess it's a symantics thing, but just because 10% run at 2.2GHz doesn't mean they run "above". You said and above. That hasn't been demonstrated. Yes I'm nit picking <G>
Otherwise I agree with your numbers.
What's that mean?
Nothing. Nothing at all.
Joe - So you disagree with my opinion that you will need a 64 bit CPU 2 years from now to earn premium CPU dollars?
No I don't disagree with you. I don't know any better than you do.
Joe - Intel will need to have either an AMD64 clone, or mainstream Itanium that outperforms AMD64 CPUs in both x86-32 or IA64 modes. Otherwise, all the premium CPU dollars will go to AMD.
I'm the first one to tell you you're entitled to your opinion but don't get too carried away and think it's anything more.
Joe - And 2 of them are 4 GB, which is the end of the road for 32 bit CPUs.
But 32bit CPUs have 36 address bits so 64GB is the current end of the road, for now.
chipguy - AC yield at x GHz = fraction of fully functional devices that pass speed path vectors at x GHz.
This makes sense. Last time you said AC yield at x GHz == binsplit for x GHz and above
Vendors Power Up Blade Server Offerings
Aug 07, 2003 (Internet.com via COMTEX) --
Hewlett-Packard Wednesday began a processor upgrade across its line of HP ProLiant blade servers and introduced the next generation of its GbE2 Interconnect Switch, a networking switch that delivers high-performance switching capabilities for HP ProLiant BL p-Class infrastructures.
Specific enhancements to the blade servers are the following: Updated the second-generation HP ProLiant BL10e blade server with a 1-GHz/1M Level 2 cache (400-MHz front side bus), ULV Pentium M processor support, and PC2100 266-MHz DDR memory support Enhanced the second-generation HP ProLiant BL20p server to contain a 3.06-GHz/1M processor with a Level 3 cache (533-MHz FSB) and Intel Xeon processor DP support Upgraded the HP ProLiant BL40p blade server with a 2.8-GHz/2M and 2.0-GHz/1M Level 3 cache (400-MHz FSB) with Xeon processor MP support
The HP ProLiant BL GbE2 Interconnect Switch is supported by a new 24-port Gigabit switch module designed by Nortel Networks specifically for the HP ProLiant BL p-Class blade servers. According to HP, the switch module is among the most advanced integrated blade system Ethernet switches available.
HP ProLiant BL architectures are designed to enable rapid deployment, provisioning, and re-provisioning of resources as demands change. The system features built-in management and virtualization capabilities to enable enterprises to adjust to demands and easily add, remove, or re-deploy applications to different blade servers and storage resources.
In recent benchmarks, these beefed-up HP ProLiant blade servers demonstrated increased performance capabilities over traditional designs. For example, the HP ProLiant BL40p blade server recently posted a Microsoft Exchange 2000 Benchmark result of 13,500 MMB2, making it the first four-processor blade to achieve this result.
(The MMB2 benchmark simulates the type of workload seen by customers as they increasingly rely on messaging services deployed in today's corporate e-mail environments.)
HP, which holds a 55 percent market share in IA-32 blade server revenue in the United States, is currently the only major vendor shipping the latest generation 3.06 533-MHz processor blade technologies. It is also the only major vendor shipping four-processor blade servers. Tatung Sharpens Its Blade Collection
Tatung Science & Technology (TSTI) this week introduced an ultra-dense blade server capable of accommodating up to 14 independent server blades in a single 3U form factor chassis.
The TUD-3114 features integrated switch blades and management blades, and supports up to 196 Intel Xeon-processor-based servers per rack. In a single 3U rack-mountable chassis, the TUD-3114 can accommodate up to 14 server blades, two switch blades, two management blades with failover capability, three redundant power supplies, and two high performance fan modules.
Each server blade can be configured with power-saving low voltage Intel Xeon processors at 2.0 GHz, up to 4 GB of ECC registered DDR266 memory (two DIMM slots), IDE 2.5 inch hard disk drives, dual 1 GB Ethernet (using Intel PRO/1000 MT Dual Port Server Connection) connected to two switch blades, video controllers with VGA and SVGA resolutions and support for USB CD-ROM and FDD .
Components are hot-swappable with each blade featuring activity indicator LEDs for the CPU, primary and secondary networks and HDD. The TUD-3114 blade server supports Windows 2000, Windows Server 2003, and Linux operating systems.
"The new servers fully integrate all the features that today's extremely cost-conscious customers demand. We have greatly increased the number of servers that can be installed in a single rack, while lowering power usage requirements using high performance Intel Xeon processors. At the same time, we have reduced the number of wire connections, and also provided powerful data management software for deployment, platform, power and workload management," said Kam Chan, president of TSTI, a U.S. subsidiary of Tatung.
Hemant Dhulla, director of enterprise volume platform marketing for Intel noted, "These servers can address the need of data centers to increase server density and manageability with high performance processors while meeting existing power requirements."
The TUD-3114 is scheduled to begin shipping later this month. The system is priced starting at $7,970 for a system with two server blades, one switch blade, and one management blade. Revario Takes the Virtual Route to Linux Server Consolidation
Revario, a vendor that specializes in open source related software, this week unveiled Revario Virtual Blade , a Linux server consolidation solution.
Revario Virtual Blade extends User Mode Linux (UML) to simplify and facilitate server consolidation. UML enables enterprises to run multiple independent versions of Linux distributions on a single server. By enhancing and extending UML with the addition of comprehensive security and automatic firewall generation, Revario believes enterprise can get maximum use in a minimal amount of time.
The result, according to the vendor, is a commercially supported, automated, and secure virtualization solution that enables enterprises to run multiple, stand-alone Linux instances on a single server in a highly configurable manner.
Revario claims its Virtual Blade product enables highly secure and manageable Linux instances totally independent of one another, as well as offering efficient, isolated, and secure development, quality assurance, and production environments. It also offers improved system and application availability.
"Revario Virtual Blade enables open source innovation to support corporate Linux server consolidation efforts," said Jason Sango, vice president of technology for Revario. "When you consider the amount of server under-utilization that exists today and the fact that the top mandate of IT operations today is to, 'do more with less,' Revario Virtual Blade strikes at the very core of what all companies are trying to achieve."
YB - Do you know what is the price of 2 Ghz part? Is it on sale now? If the price is less than twice of 1.8 Ghz part then AMD is clearly getting close to 15% binsplits.
For the sake of discussion let's assume you're right. This doesn't tell you anything about yields though. They could have 30% yield with excellent binsplits.
One more point, we know that IBM is consuming at least 2,000 2 Ghz parts this quarter. So, AMD must have at least 5,000 chips coming. They must bin better than 5% to collect this amount.
How many wafers do they have to run to get those 5K parts?
chipguy AC yield at x GHz == binsplit for x GHz and above
Run that one by me again???