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.
Elmer,
I mean reason, AMD has never produced at anywheres near capacity, even when the market was willing to buy.
Well, I expected a lot more than about 7 million units when AMD said they were capacity constrained, and Austin was still making CPUs and Dresden was ramping up.
The highest number of CPUs "sold" was 8 million in Q1 2002, when AMD was producing only in Dresden, and the CPU was a 129 mm^2 Palomino. That is not horrible IMO, and AMD did not specify capacity constraint at that time.
Joe
Elmer,
Sorry but it didn't add up. Simply taking the die size showed that AM could have made a killing if the simply produced at anywheres near capacity, assuming normal yield. AMD didn't deliver and the excuses were legend. The only one that didn't require yield problems was that AMD simply didn't think to start enough wafers.
Not that they "didn't think" of starting more wafers. It is logical to limit risk. If desktop Palomino turned out to be buggy, AMD would be out of business.
But let me ask you a related question. When did Intel go from one core to entirely different core within 1 quarter? P4 ramp ramp took 4 quarters.
Joe
Elmer,
In Q4 of '01 Intel was capacity constrained by the oversized Willamette die. AMD was in a position to sell anything they could make. AMD delivered about 1/2 what their capacity should have been able to deliver.
The same way Intel was capacity constrained on Willamette, AMD might have been constrained on desktop Palimino (there a grand total of about 1 reference to the fact from The Inq). You may recall that this was the first quarter of sales of desktop Palomino, so capacity constrained on a brand new CPU could be understandable.
At the same time, Intel was drowning in P3 die, and AMD was drowning in Tbird die, so there was no capacity constraint, just Intel moving the market away from Piii and Tbird to greener pastures.
On top of this supposed shortage, there was a serious channel stuffing (which is a contradiction), that kind of blew up following Q1 for AMD.
Anyway, I feel pretty confident standing by my statement that the last time AMD was capacity constrained was Q3 2000.
Joe
sgolds,
My understanding of PAE is that it is just an API. It is not a change in the 32 bitness of the program. Nor it is a change in the fact that this program can address only 4 GB of data. THe API just cleverly shifts where the CPU is physically looking while addressing the same logical 4 GB of memory.
So there is no recompilation or anything like that. Recompilation, wile making changes to how pointers to data are handled is IMO extremely tricky, and it could be a recipe for a serious disaster.
Any tweaks in addressing capability of Prescott IMO would needs to be considered a brand new instruction set, not unlike x86-64 being a brand new instruction set.
Joe
sgolds,
Isn't there a ton of code out there that does some weird stuff with addresses, computation of offsets etc. where the compiler has no idea that the data being manipulated is an address until the very last moment where it is used to reference data?
Joe
Paul,
But I think that with the products AMD has coming, there's going to be plenty of demand, at higher ASPs than they'll get from Dell, to use up their capacity.
Don't be so sure. The last time I recall AMD to be production constrained was Q3 of 2000. I for one would welcome a day when AMD has to say to a $35 CPU buyer that AMD is sold out, and that if he wants a CPU, he will have to double his bid price next time or go to competition.
While I am somewhat optimistic about Opteron, by no means do I take it for granted that AMD will be able to sell out the production any time soon.
Anyway, back to Dell, with Dell aboard, the ASPs would shift up, and the bottom customers would just fall off, based on AMD's production capacity.
Joe
Not a Short,
There are some benchmarks for products, others for companies. In benchmark for companies, Dell is certainly at or near the top in most categories.
About their products, well, I haven't bought from them since they were PC Limited, but there are millions and millions of people who buy their stuff, and I am not going to argue with what works (for all parties).
As I said, it will be a great day for AMD shareholders when Dell comes aboard as a customer.
Joe
wbmw,
IBM might lead in server revenue, but HP sells in higher volume.
I could agree with that.
Joe
Greg,
That reply was unnecessary.
Hmmm... Being about 400 posts behind, I didn't know that others already said what I said, but much more tactfully.
BTW, I noticed that the moderators removed my post. This is the very first time ever that has happened to me, so I guess I need to be more carefull.
Joe
wbmw,
2. Microsoft has no reason to invest in Opteron optimizations. That was the original subject for #2, was it not? It sounds like their support will be minimal, and highly leveraged off of the work they already did for Intel.
I think you are mixing 2 subjects: SpecInt optimization and general optimization to perform well on current hardware.
Microsoft has no reason to invest in SpecInt optimizations, but I expect a competent C++ compiler and .NET Opteron engine. Microsoft wants to sell software, and has interest in Microsoft software performing well against the competition. Microsoft software is compiled using Microsoft compilers.
Microsoft has already done some of work on Opteron optimizations by working on 64 bit Opteron compiler and by taking advantage of the extra registers.
5. Windows 2003 Server is missing the .NET framework for Itanium 2. This is unfortunate, but not a barrier for entry right now.
If you are considering running any .NET code on the server (ASP.NET, ADO.NET or custom apps) it is an insurmountable barrier for entry.
Itanium's competitors will not be using .NET, and it's doubtful that Microsoft will get a 64-bit version for Opteron before Itanium.
Opteron can run .NET as is. 64-bit version will be an icing on the cake.
Joe
Fyo,
Another datapoint in the power consumption discussion is the head Newisys guy, who said that the actual power consumption of Opteron is in around 40 Watts (or 40s).
If anyone should know, this is the guy.
Joe
wbmw,
On the other hand, AMD developed interesting MirrorBit flash designs, yet AMD's technical documents suggest that they are all still at 230nm
That was a good find on your part. I think you are the only one who looked this up, while most people assumed that Mirrorbit exists on 170nm.
BTW, 130nm manufacturing has begun. I don't know about Q2 revenue, but Q3 should see a strong impact from 130nm Mirrorbit flash, since it does not look there will be many competitors in the densities AMD is talking about. It seems that only Intel with the massive R&D budget can pull a rabbit out of the hat to keep up with AMD here. STM seems to be throwing in the towel on NOR flash.
Joe
Elmer,
I have no problem with different standards. Not for AMD and Intel investors, but for posters and hecklers.
Joe
Elmer,
I quickly looked around and found following in stock
http://www.insightcomponents.com/ic/apps/tinas/index.php?page=48
240: 91
242: 55
244: Ordered upon request
http://www.yesmicro.com/shop/product.aspx?pid=6163153&seid=4
240: 31
242: 47
244: Backorder
http://www.shopperwiz.com/shopdisplaydetail.asp?REF=D_TOP_SEARCH&id=840010
240: 40
242: 8
244: Backorder
http://www.techdepot.com/Product.asp?ProductID=1638560&iid=342
240: 81
242: 55
244: Backorder
http://www.shopblt.com/cgi-bin/shop/shop.cgi?ACTION=thispage&thispage=011003002001_B840010P.shtm...!
240: 70
242: 16
244: ETA 5/8/2003
Only this vendor is listing 244 in stock: http://www.pcprogress.com/product.asp?m1=pw&pid=AMDOSA244CC05AH
I think this is a good availability for something that is supposed to be a server CPU, one week from release, somethng that is taking Itanium 2 years to achieve.
Of course Xeon availability is far higher, 20 fold if not more, but Opteron may get there one day.
I'm still waiting to see if they can deliver meaningful volume
Well, as usual, I am waiting to see if AMD can sell meaningful volume.
Joe
chipguy,
The subject is whether or not IBM can supplement AMD production capacity of 130 nm x86-64 devices.
I don't think anybody is talking about IBM supplementing AMD production capacity at 130 nm. If the current process development partnership continues, 65nm will be the earliest time frame.
Joe
Elmer,
The issue was why is Opteron's die so large compared to Athlon
Well, is it large? Forget this Subzero character and consider the info that's available.
Fyo took some measurements in this post:
http://www.siliconinvestor.com/stocktalk/msg.gsp?msgid=18871892
Rounding the memory controller to 5% I get that the core (including L1) at 25% of the core, L2 + memory controller + memory interface + HT interface at 75%.
Starting from 193 mm^2, the core is 25%, or 48.25 mm^2. Now add 512K L2 (1/2 of 42 or 21%) of the die (all based on Fyo's figures) which is 40.53. These 2 figures add up to 88.78 mm^2. Let's compare this to Barton, which is over 100 mm^2, so Opteron is actually smaller. Comparing apples to apples, we need to add EV6 measurement to Opteron.
But the bottom line is that the statement "Opteron's die large compared to Athlon" is not true.
and additionally why is it slower
Why is the clock speed lower? Good question.
Joe
Paul,
I, for one, have never been real excited about the possibility of Dell for AMD. They'd have to be a lot bigger to be anything more than solely Dell's slave, and the margins would likely be less than what the rest of the market will provide.
Margins, at least for AMD are much more associated with the volume and ability to sell out its production than with ASP being $5 higher or lower. Dell would provide AMD with steady demand, increased volume, and therefore _higher_ margins.
Joe
sgolds,
I agree, a flat address space Prescott with 32-bit logic registers would be a sensible stop-gap product for Intel.
Wouldn't it create as many problems as it would solve? It would need to have a new instruction set, that is neither 32 bit nor 64 bit. It's kind of like deciding to switch from driving on the left side of the street to the right site, but making the transition gradual.
Joe
wbmw,
I read somewhere that IBM is #1 server vendor, but I guess it depends on how you count. A mainframe can be considered to be a server, and that is how IBM probably becomes #1.
Joe
yb,
Dell is a joke, screwdriver shop - they can't make any complicated hardware.
Dell is a great company, IMO, and I would like it to become AMD's customer.
Joe
Elmer,
Now tell me, was this a vapor launch?
I think it may turn out to be a good launch, similar to Barton. Barton started with only 2800 speed grade available first week, then 3000 week later, and 2500 2 weeks after launch.
I think the 240 and 242 parts are available already, I haven't seen 244 as a Boxed part yet, but OEMs seem to have it available in their boxes. The 244 part was supposed to become available in May, so we will see. If the 244 is in short supply or unavailable within next 4 weeks, I would say that it was below average launch, if it is available, I would say that it was a good launch. Not a vapor launch.
Joe
Elmer,
Although we are all impressed by the excellent benchmarks Opteron posts, there are still many unknowns.
I think some of these unknowns will be answered in next 2-3 months, when Opteron makes it to workstations, and can be put through paces by benchmarkers and overclockers. Also, in this time frame, if there is a new speed grade, it would indicate that things are on the right path, if there is none, it would be a warning flag.
Joe
smooth,
What is IBM's manufacturing capacity, any more than Dresden? They can't be making money on their fab(s).
Obvoiusly AMD will use Dresden before outside foundry. But AMD can sell more than Dresden can make, IBM foundry deal would be a way to increase production. It'ssomething very hypothetical, since it has been probably 3 years since the last time AMD was capacity constrained.
Joe
subzero,
You are jumping to way too many coclusions:
A SLOWER Opteron Processor with SOI than the Athlon without SOI - 1.8 GHz Oppie vs 2.1+ Ghz (Athlon)
If AMD never releases another speed grade, you will be correct. What do you thing the chances of that are?
a whopping 85 watts for the 1.8 GHz Opteron vs. less than 60 watts for a 1.8 GHz Athlon.
Since it is not possible that 1.4 GHz chips would dissipate as much power as 1.8 GHz, the 85 Watt number is clearly wrong for actual power consumption.
The Oppie with SOI is a horrendous 193 sq. mm. - where AMD had promise 104 sq. mm. several years ago.
AMD never promissed 104 mm^2 to be the die size of Opteron.
1.8 GHz - 193 sq. mm. - 85 watts - SOI is an abomination.
A conclusion supported by 3 wrong assumptions ...
Joe
econo,
I agree with all the points in your long post to sgolds, but I just don't follow your argument in this post (I am replying to). And by the way, this is about 3rd time you presented this theory (the other 2 times on RB) and I didn't reply there, not because of arrogance or lack of respect (I understand you put a lot of time writing those posts), but because I just didn't understand your argument.
IMO, AMD has a huge sales problem that is unrelated to capacity to manufacture product (well maybe tangentially in periods such as mid 2002 when AMD wasn't able to produce higher value products).
When there is performance parity with no shortages, AMD needs to sell its products for $60 to $90 in order to make sales, while the competitor is able to sell the same product for $120 to $180. Sometimes, such as Q1 2003, AMD sells chips with market value of $150 for $60 (downbinning), because AMD is able to make a sale on this chip when price is $60, otherwise, the chip would go unsold. And even at $60, AMD does not sell all it can make.
At the same time Intel is able to sell its $150 chips for $150 - for number of reasons - nice or not so nice.
If they could sell 12 million desktop and mobile chips per quarter they would break even with ASP around $50. Selling 6 million they need over $80 ASP. At the higher prices it is much harder to get even the smaller volume of sales. It is largely AMDs ability to make enough chips that determines how aggressively they can price.
You sell what you can sell. To some a small extend, the chips are a commodity and they sell themselves if the price is low enough, in a certain market segment.
What I mean by selling is agreing with a customer (OEM) that you will supply him with microprocessors for a line of his product, that you will sell them at some predetermined price (influenced by market conditions going forward), and once you have made this sale, you have a built in demand for your chips from this OEM as long as he offers this line, as long as this OEM can sell the products successfully to the end users. This is where AMD is lacking the most. Of course there are good reasons (competing with a near monopoly with Bath party mentality), but AMD achieves some successes, there just need to be more successes.
The other part of the market is the fair weather friends, such as white box, some customers in third world, some DIY, who just always want rock bottom price. This is basically the commodity segment of the market. An unusually high percentage of AMD sales seems to go to this category rather the former. It could be that these people just want the rock bottom price, and Intel can't force them to buy Intel, or punish them for not buying Intel. So AMD continues to be in this limbo, selling to these outlets, and having to sell low priced product, just because even a $40 sale is better than no sale at all.
Now suppose AMD can make 40 million chips per quarter tomorrow (100% of the market). How is it going to change things?
At 5-6 milliion per quarter they need to maintain ASPs in order to survive, and they have too few chips to really benefit from aggressive pricing to gain share. With more output they can price as low as necessary to clear output.
I am completely lost. AMD already has problems to clear output, and way they clear output is that they sell silicon that has $150 market value and they clear it for $40. Sometimes even that can't be achieved.
I guess what you may be thinking (I am guessing) is that AMD produces a lot of chips, the costs will go down, and AMD can just flood this commodity segment with chips, sell them for $40 or $50, and completely survive and prosper just in this commodity segment, ignoring the quality or status conscious segment that spends $150 for CPUs.
Well, this commodity segment doesn't have infinite demand. You would still need to gain share from the $150 segment. But it can't really be done by saying that you have a $40 CPU, and they will come. AMD already does that, and the $150 crowd rather pays $150 (and largely pays it to Intel rather than AMD), in order to maintain their perception of quality or status. I just don't think this is the answer for AMD.
AMD needs to sell to this $150 crowd, and to sell to this crowd for $150. To do that, AMD needs to offer quality / status that would go along with that. Competing with Intel, makes things so much harder, but AMD was on its way there with Athlon, before the setbacks caused by delays of Palomino and Tbred B. The launch of Opteron is a new start to regain this status. I am a shareholder because of Opteron and A64, and maybe even potential of Barton to regain this status.
Let me go back to Q1. AMD clearly could have produced a number of chips far in excess of 6M units. Theoretically, AMD could have produced 8M, 10M or if fully utilized on Tbrd B, even 13M chips. But forget these large numbers, and let's say AMD could have (or in fact did produce) 6.5M units. Or let's put it differently, had available for sale 6.5M units in mix of Tbreds and Bartons, yet they sold 6M. (Well, I don't know if it was 6, but let's say that unit supply of chips for sale was 0.5M higher than actual sales). How exactly is this higher output benefiting AMD in their quest for what you mentioned in:
With more output they can price as low as necessary to clear output.
?
Joe
sgolds,
Didn't the deal include jointly building a 65nm 300mm factory? Not hard to do a little joint manufacturing with that start, is it?
The original plan was to build the fab jointly with UMC, but that seems to be on hold. As far a joint manufacturing, that has not been announced, only process development collaboration. It would be an easy next step from there to manufacturing, but manufacturing in the existing IBM fab in East Fishkill.
Joe
yb
I think the consensus is that ASP in Q4 was around $65, and remained the same in Q1.
Joe
yb,
re: last Q ASP
There was some talk of dumping a large number of cheap chips in Chinese market, which may be the reason for no better ASPs, but the main reason is that AMD didn't sell enough of the higher priced Tbreds and Bartons.
Joe
sgolds,
don't forget that large numbers of wafers will be needed to build inventory for Athlon64 for end of year sales.
There is no reason to package any A64s if you are selling out Opterons. If AMD keeps selling out Opterons, A64 can be a low volume launch or postponed launch.
There is just no reason to be supply constrained on Opteron (unless the SOI just completely doesn't work). So basically, AMD first allocates wafers fir all the high end Opteron chips that can be sold and goes down from there.
I don't have any numbers to make a reasonable estimate of how many wafers will be needed for each processor but I am having a hard time believing that AMD will be able to produce enough Opterons this year to be market constrained.
Well, I gave you one ballpark number. Here is another. AMD allocates half its Dresden capacity 2500 wpw to Opteron, achieves 60% yield, or about 90 good die per wafer and produces 2.9 million Opterons per quarter, and achieves 100% market share and on top of that places 400,000 to inventory.
Now think about which one is within realm of possibilities, ability to produce this many chips or ability to sell this many chips.
If the demand is there for Opteron it makes no sense to forgo selling 1 $500 Opteron in favor of 2 or 3 $60 chips. There is no reason for Hector to leave revenue of $1.25 billion per quarter, even if it took 100% of the wafer starts to produce these 2.5M Opterons per quarter.
The obvious answer is that AMD is (and has been for long time) demand limited. Even at prices lower than Intel, AMD has hard time selling out the products, and Opteron is no different.
The reason why there are no more OEMs selling Opteron is not supply (well, some may be waiting to see proof that AMD can produce them, since it is a brand ne chip) but other reasons.
Joe
chipguy,
And with MB levels of on-chip cache by the time you get to the uP pins most of the memory traffic for most
programs is almost entirely data so code traffic is seldom an important factor.
i was kind of wondering about the ratio of code / data in cache, but there are probably just too many variables here to come to any reasonable estimate. But as the cache sizes increase, you are right the data begins to predominate (running the same set of programs).
I guess this is the reason for such large caches on Itanium, since Itanium executable proably have the highest memory footprint of any current processor. So, Itanium has to first get over this code hurdle before data can be cached.
Joe
sgolds,
I suspect Sun wasn't there yesterday because AMD could not provide Opterons for them this year. IBM most likely has the contract for the bulk of this year's production, with scraps going to the white box makers./i>
I highly doubt it. AMD is ramping SOI wafer starts, and shortly, AMD will have capacity to supply more than 100% of all servers sold. This supply problem is generally just a product of Elmer's imagination. AMD's constant problem is demand, or ability to sell what they can produce.
Most widely used figure for server CPU shipments is 10M per year, or 2.5M per quarter. If Opteron manages to get 10% of that, it is 250,000 per quarter. Even at fairly low yield of 50 good die per quarter, all AMD would need is less than 400 wafers per week to produce them.
I don't think AMD will reach 10% in 1st or 2nd quarter of Opteron availability.
Joe
You may want to add MicronPC to the list yb posted. Their systems have nVidia nForce2 chipset, which is the best pefrorming chipset for Athlons.
http://www.buympc.com/smallbiz/store/desktops/product_detail/millennia_910a.html
I am sure other higher end vendors such as Alienware also offer nForce2 based systems, but I believe current HP crop is based on integrated Via chipset (not as good as nVidia nForce).
Joe
It looks like Opteron 242 is in stock at some vendors:
http://www.computers4sure.com/Product.asp?ProductID=1638561&iid=342
http://www.ebuyer.com/customer/products/index.html?action=c2hvd19wcm9kdWN0X292ZXJ2aWV3&product_u...
http://triointernational.com/Store/view_cart.cfm?shipwzipcode=&prod_num=1003%20%20%20%20522644&a...
http://www.truedataonline.com/computer/resultsdata.asp?part_no=30840013
I got this info from here:
http://amdzone.pricegrabber.com/search_getprod.php/masterid=773724/search=opteron/ut=d82e4de28cb61b6...
Joe
yb,
I think workstations will be a killer once a good fully loaded dual-cpu atx mainboard will cost under $200 and 1.6 Ghz Opteron will cost $300. This way you will get the nice setup.
I think Opteron will need DDR400 support using unbuffered memory, and clock speed of 2 GHz.
Joe
Combjelly,
AMD is being rather coy about these numbers, aren't they? I wonder why?
One possibility, or something I would hope to be the answer is that the power consumption (as well as voltage and frequency) depends on load, the way PowerNow operates in notebooks. The differences between the speedgrades would then be just that 1.8 GHz chip has 9x multiplier available to it, while 1.6 would be limited to 8x etc.
But even under this scenario, it should be possible to specify power dissipation of a processor running at 1.8 GHz using voltage for 1.8 GHz running some program that puts maximum stress on the CPU non-stop. This number should be different from 1.4 GHz CPU which can possibly be using different voltage...
Joe
spokenshave,
Opteron tech doc does not have that column filled in. "Typical" would roughly correspond to Intel's use of Thermal Design Power (TDP)and, if Athlon is any guide, "typical" power will be about 22% less than max power. So, look for the typical power for Opteron to be about 66W.
There is just no info to go by on the power dissipation of the actual chips. The Newisys guy said it is around 40 Watts, but I don't know if this is under full load, idle etc.
The voltages make no sense either. Hopefully some info will emerge.
Joe
Elmer,
The last time I checked, some online vendors had ETA of 4/25, which is end of this week. But it would be a better idea to check with the white box vendors selling actual systems (rather than processors alone).
I know there was a post that listed all the Opteron vendors. Can someone repost the link? I would be curious about their availability.
Joe
Elmer,
I have been looking through the review and specs, but I haven't found any definitive answer. All I found was the overall maximum TDP regardless of frequency. There is no info per clock speed. The info on voltages is just as sparse. I only see 1.55V spec, but that is the absolute highest voltage for Opteron, so that doesn't make sense either.
Joe
wbmw,
I think it is a great idea to use laptop chip for servers, or, a chip that supports laptop-like features, especially the adaptive clock speed and frequency adjustments. I think there is a great percentage of servers out there that are idle 99% of time and have short bursts of activity 1% of the time. This would provide excellent energy savings, and, assuming a number of servers in the server rooms have this functionality, he overall cooling, powering could be reduced.
I like the environment angle of this. It shows that with human ingenuity, you can have _more_ computing power for _less_ electric power. (Unlike the environmentalist activists who keep talking about what we can't or shouldn't do (or have) to help the environment).
Of course process shrinks do that as well, but this is on top of the shrinks, and it is more dramatic.
Joe
wbmw,
Meanwhile, in three years, Intel will redesign Xeon based on the Centrino (Pentium M) core.
I am not sure if you are kidding or not, but I think it would be a good idea. There is no reason why it should take 3 years (unless you are talking about adding 64 bit support)
Joe