Slide Show

Add Your Comments

Yesterday, I mentioned that AMD had an analyst conference with with lots and lots of slides

Today, let’s look at more of them, and see what they tell us.

For The Short and Medium-Term

20062007

This confirms much of what we already knew about the 2006 (please note if you haven’t already that “Presidio” is AMD’s version of TPM, which is the hardware DRM you’ve heard about on the new Intel boards), but adds a few interesting tidbits:

  • What is “AeroGlass graphics?” One might think it’s a fancy name for integrated graphics, but it’s also listed in the “performance” column, so we’ll just have to keep that in mind for the future.
  • DDR2 will have a short lifespan on the AMD platform, in 2006, replaced by DDR3 in 2007. This is especially important given the onboard memory controller on AMD processors, which so far have been designed to handle one memory type only.
  • AMD also promises a “new core” in 2007, though from other indications, calling it Hammer II would be more accurate. It promises bigger caches, DDR3 compatibility, a faster Hypertransport, and PCIe II.

    Don’t Expect Big Speed Bumps

    This shouldn’t be a surprise (and this applies just as much to Intel’s next generation), but for those on a hope high:

    Power

    Going to 65nm will help matters a bit, but for at least the next few years, performance will compete against power consumption, and performance is going to usually come in second. Multicores alone will make sure that’s the case most of the time.

    Tinkering Away

    Here’s an interesting item for those who find current Hammer not quite good enough:

    20062007

    We’ve long known both CPU companies tinker away with their products, but this is the first time AMD’s actually described theirs in any detail.

    AMD is supposed to be coming out with an FX-57 shortly which is supposed to run at 2.8GHz default. Maybe that’s the next “upgrade,” one which should trickle down the next few months.

    Deeper Pipelines and More Far Future Stuff

    Here’s something we haven’t heard about before:

    Pipeline

    It seems clear AMD plans to eventually deepen their pipeline, and they think they can get more performance from doing so. AMD has more leeway than Intel here, and having a deeper pipeline could help heat issues.

    Perhaps more importantly, having a processor that architecturally resembles the PIV means that CPU speed will be more comparable than they are today. Irrelevant in fact, but important for Joe Sixpack. Looking faster is often more important than being faster.

    We should note, though, that this at very best is a 2007 development, and most likely post-2007. So don’t hold your breath for this.

    Don’t hold your breath for the items mentioned in column #3, either:

    FarFuture

    The “on-chip coprocessors” sound awfully Cell-like, but again, this almost certainly is post-2007.

    Look in the middle column, though, and you see reference to FBDIMMs. For those interested in that, looks like that will become an option for at least servers next year.

    Tangential Tidbits

    MultiYears For Multithreading

    For those who think multithreaded applications and games are going to start popping up like weeds, this AMD slide should (quite rightly) burst that bubble:

    ParallelProcessing

    (If you don’t know what Amdahl’s Law is, look here. If one click is just too much work, AMD is saying that you can’t normally expect a ton of improvement from parallel processing simply becauwe many normal computing tasks just aren’t parallel in nature.)

    That’s correct, and a good point to make, but ever since this began, AMD just doesn’t seem too happy with the thought of duallies for the desktop, certainly much less than Intel. This is pretty odd for a company that is promising dual-core across the board in a year; it’s like they’re doing it because they have to, and they’re sulking about that.

    Frankly, “understanding the appropriate rate of transition is what will be important to real customers” sounds like an excuse for “not keeping up with Intel.” Frankly, when AMD doesn’t like a new technology, or doesn’t think it’s all that important, that’s because they can’t make it yet, or make it well.

    However, you don’t need multithreading to get advantages from at least dual cores, you just need a customer who does (or would like to do) more than one serious thing at a time with his or her PC.

    The comments about multithreading are accurate, but there seems to be more here than just that. I think single-cores are going to play a much bigger role for AMD the next couple years than the roadmaps indicates.

    Keeping The Penguin Afloat

    AMD makes a statement you don’t run across very often when it comes to Linux:

    Linux

    There’s nothing wrong with AMD doing this; it just makes you wonder about Linux a bit.

    The image of Linux is often that of an army of incredibly skilled, incredibly selfless programmers doing this solely out of the kindness of their hearts: Sisters of Computer Charity, if you like.

    You see items like this while remembering all the Linux programming IBM does, and you have to really wonder just how much of Linux (outside of specific items like device drivers) is really written by cyber-Mother Teresas, and how much is written by very paid programmers working for very commercial firms.

    Volunteer work is one thing, donating work in an act of enlightened self-interest is just about as good, but taking credit for someone else’s heavy lifting and bill-paying, well, that’s not so good. The heavy lifting and bill-paying is fine; it’s taking the undeserved credit that’s the problem.

    SOI: Half-A-Loaf

    SOI1

    This graph is entitled “SOI: Dramatically Improved Performance and Power Efficiencies” and goes on to say that SOI “provides up to 18% performance improvement” over plain vanilla silicon.

    However, up to 18% is far less impressive for those with sufficient memory storage to recall that IBM promised double that, or up to 35%.

    Why is that important? That’s why people aren’t hitting 3GHz on air with Hammers. That’s why IBM didn’t hit 3GHz on the PowerPC and couldn’t quite deliver a mobile G5, which gave Apple and Macsters at least a good excuse to switch to Intel. This is why AMD’s Turion isn’t quite power-competitive and thus not too popular in notebooks compared to Dothan.

    This little statistic punches above its weight.

    More To Come

    On Friday, we’ll present part II of this slide survey. Part II will point out the cute little tricks a company can play in a presentation like this.

    Killing A Few Birds With One Stone

    Today, we were supposed to speak about presentation tricks, but there was one item that didn’t precisely fit the mold: production capacity.

    Outside of the slides, AMD has been making some recent noises about this.

    Near-term, you had Hector Ruiz telling an interviewer “There’s no reason why we shouldn’t aspire be a third of the market in the next two or three years.”

    Far down the road, you had talk in a newsletter about AMD reaching 50% marketshare by 2015.

    In the last couple days, the Frankfurter Allgemeine Zeitung reported that AMD is in talks with the Germans to help finance a third Dresden fab.

    Finally, we have this nice little graph estimating AMD’s production capacity over the next few years:

    Capacity

    What are we to make of all this?

    Looking At the Graph

    Before we look at the numbers year-by-year, two points:

    First, the graph gives the first real indication how much third-parties like Chartered can be expected to increase production.

    Second, theoretical increases in capacity due to die shrinks have always been tempered by increases in CPU cache, and this will be especially so over the next few years. Average cache size will double for Intel chips, and the same will probably be true for AMD chips, too, so much of the potential increase in CPU production will be eaten up by this.

    Third, dual-cores obviously change the number of CPUs that can be made by a die a lot. Whether you forge them from a single die or just glue them together, a dual-core will chew up twice the fabrication capacity of a single core chip with the same cache. 90nm dual-cores are huge by single-core standards. A 65nm shrink will help a lot, but a 65nm dual-core will still be a bit bigger than a 90nm single. They’ll be tolerably bigger, but if they become the mainstream chip at 65nm, there will not be the usual capacity increase due to the shrink. bonus

    Having said that, the graph indicates that AMD’s die output will go up quite a bit in the next few years:

    2004: 30 million
    2005: 47 million
    2006: 57 million + up to 10 million from Chartered
    2007: 63 million + up to 22 million from Chartered
    2008: 72 million + up to 26 million from Chartered

    (Incidentally, the chart isn’t drawn to scale; the distance between 50-100 million is about 8% shorter than that between 0-50 million. This makes Chartered’s contribution seem smaller to the eye than it actually is.)

    For those claiming that AMD can’t make a $300 dual core due to capacity restraints, well, there seems to be plenty of capacity opening up in 2005. 🙂

    Looking further down the road, these numbers are very hard to judge for the reasons mentioned above, though it does seem like AMD’s production really doesn’t go up all that much in 2007-2008 even after Fab 36 is up and running. The numbers seem too low for two fabs making CPUs full blast; a little too high for just Fab 36 running full blast.

    AMD has never been all too clear on what it is going to do with the first Dresden fab after the second opens. On the one hand, they’ve indicate that there would be some updating of Fab 30. On the other hand, a good part of this presentation talks about AMD’s initiatives in “pervasive computing” (read: embedded chips). Maybe Dresden will largely make older chips to address that market.

    The Likely Need

    In 2004, about 190 million PCs were sold (Gartner).

    The number of PCs made the last few years has generally grown at a rate of between 10-15% yearly, much less in North America, rather more in Asia.

    For the next few years, it’s probably safe to say that PC demand isn’t going to get much better than that (and there’s more reason to think it will be worse, more below).

    Assuming no serious downturn, if we assume a PC growth rate of 12% yearly, we get:

    2005: 202 million
    2006: 226 million
    2007: 253 million
    2008: 284 million

    If you look at AMD’s capacity figures (and assume that they can get about the same number of CPUs per million capacity as they do today); those figures are just around the 33% market share figure.

    After 2008, you can see why Hector wants a new fab.

    Debeautifiers

    Over the next few years, two factors could make the PC industry a good deal uglier and AMD’s job a lot harder.

    First, it is very questionable the next three years are going to be OK-or-better years for the PC industry. European economies have slowed down a not-too-fast pace, and the United States can’t keep borrowing its way to a better time forever. Eventually, the US will have to slow down its spending one way or another, and, more likely than not, that will put a huge damper on the rest of the world.

    Recession and slow/no PC growth will only increase the size of the CPU glut, and make many look for cheaper options.

    Hello, consoles.

    It’s unlikely the consoles we’ll see in 2006 will make an immediate big difference in PC sales. They may well not end up affecting them much at all, and even if they do, it will take years to them to have a big impact.

    Nonetheless, any little bit hurts, and more importantly, if consoles like the PS3/XBox 360 look like they’re going to take a big chunk out of the PC market in 2007-2008 (and the people who make them/back them do have lots of money), well, no mainstream CPU company is going to look too attractive, and the weakest one will look downright ugly.

    The Real Problem

    Up to now, Hector’s actions have (quite understandably) focused on the goal of making more money from the CPUs AMD makes. AMD still isn’t quite there, but if Hammers can become AMD’s standard chip and more-or-less keep pricing where it is, AMD could start making a couple hundred million dollars a quarter from CPUs in a few quarters.

    However, after that, all these plans assume AMD can add 15% excess capacity to the CPU market, rip that away from 15% marketshare from Intel, and do this without affecting CPU prices too much. AMD is going to deliberately create an excess of supply over demand, and Intel is supposed to just sit there and shrink?

    Let’s assume the very best for AMD, that if they make it, people will come. Assume they’ll have no further problems making chips. Assume people will know about them, assume people will want them over Intel, assume people will pay Intel-like prices for them.

    Let’s assume everything goes AMD’s way except the part about Intel rolling over and playing dead. What would happen?

    Intel will never, ever let AMD grab that kind of marketshare without a big fight, and when all else fails, what does a highly profitable company with lots of spare cash do to a far less profitable, highly-indebted competitor?

    Can you say price war?

    In the AMD-Intel competition, if push ever really comes to shove, Intel will always win simply because they have money and AMD doesn’t. Period.

    Intel can always turn any serious threat by AMD (that means seriously breaking out of the 15-20% marketshare ghetto they’re usually in) into a financial game of chicken, a fiscal death match, if you will, and AMD will always run out of money before Intel. This is the Intel nuclear option, and AMD cannot, by itself, ever defuse or deter it.

    For AMD, lack of money is the root of all evil.

    The only way AMD can defuse/deter that bomb is to become absorbed into a company that a) does have money and b) is willing, eager and able to use it against Intel.

    Against AMD alone, the threat of a price war is quite credible because there would be no doubt as to who would win. Intel’s profits would be hurt for a while, but they can afford that. If AMD became part of, say, Microsoft, though, now both sides have nukes, and the results of a war look more like mutually assured destruction.

    It’s not too hard to find tech companies with a), it’s item b) that’s the problem. Nobody want to play fairy godmother to AMD, probably because a second big heavy in the CPU market because any newcomer would end up shelling out Intel-like expenses in return for much less than current Intel-like profits.

    You have to wonder whether Mr. Ruiz has it in the back of his head to sex up AMD, make it more desirable, then sell it to one of the big boys who can pay the big bills.

    I doubt he does, but he should, though future events and circumstances could toss mud on his babe’s face.

    We’ll get to the cutesy presentation tricks tomorrow.

    Games People Play

    Sometimes you can talk a lot without saying anything.

    This works for graphs, too.

    There’s a troubling number of graphs in the AMD presentation that are partially-to-completely useless and/or misleading because they don’t tell you something, but rather try to make it look like they’re telling you something.

    This happens far too often for it to be accidental. This becomes even more apparent if you’ve been following the company and realize the graphs in questions cover areas AMD usually won’t talk about when asked.

    These slides are being used at an analyst conference. The people who go to these things are not kids who run websites. These are grown men and women with advanced degrees who not only get paid a lot of money, but at least in theory are getting that money to think.

    As you’ll see, figuring out what’s wrong with the following graphs is not rocket science, all it takes is some observation and an unwillingness to just swallow everything you’re told. This is certainly not too much to ask from people getting paid big bucks.

    Today, we’ll look at some of the tricks AMD plays at events like analyst conferences to make reality, well, better than reality.

    The Obligatory Warm And Fuzzy “We Love You” Slide

    Love

    Admittedly, this is commonplace, but I never cease to be amazed by corporate America apparent need to tell their customers that they deeply care for them, if not downright love them.

    How can any company love you when they don’t even know your name?

    Maybe I’m strange and old-fashioned, but if you actually like (never mind love) someone, you’re honest with them. You don’t try to mislead or misguide them or withhold information they need to know from them.

    This slide uses AMD’s favorite buzzword, “Customer-centric innovation.” If you’ve listened to these conferences, this is the go-to buzzphase used when all else fails.

    Yet AMD never tells anyone just what “customer-centric innovations” they’ve ever made.

    Another warm-and-fuzzy term used in the slide is “Expand all connections with customers, partners and end users.” That’s nice, but on the whole, the motto AMD seems to go by is, “Top Secret.”

    Just for one example, just how “customer-centric” is it when you stop updating your technical specs and demand an NDA before they’ll let someone look at them? Or repeatedly refuse to tell stock market analysts how many CPUs you’ve made, or what price you got for them (even though they can still estimate based on other sources)?

    Sounds like “Do what I say, not what I do.”

    Cite Insignficance

    One way to look better than you actually are is to cite a statistic likely to be garbled in your favor.

    Love

    The first example is “In Q1, more than 55 of Forbes Global 100 companies or their affiliates used AMD64 processor-based systems to run critical enterprise applications.”

    That sounds like AMD is taking over, but what does it really say? It says that 55 out of the 100 bought at least one Opteron system. That’s all. If more than half of the Forbes 100 had decided to switch from Intel to AMD, AMD would have a lot more than a 5.7% market share.

    The slide also shows a graph showing that AMD holds a 27% market share in x86 4-way servers. That sounds impressive unless you know that 4-way servers are only a tiny part of the overall server market (the vast majority of servers have just one-two CPUs), and that AMD’s overall server share is only about a fifth of that 27%.

    Math Is Hard

    3-way

    This graph sure makes it look like AMD has been growing awfully fast, more than 40% if you just look at the relative size of the bars.

    But did you notice there’s no numbers on the side of the chart, or any explanation as to what is being measured?

    It’s not CPU production drawn to scale, that didn’t go up 40%. It’s not marketshare drawn to scale, either, for the same reason.

    There’s a very good reason why there’s no numbers on the side of the chart. I think this chart appears to represent AMD’s market share, with the bottom of the scale set at 12%. This chart fits the data we get from third-parties pretty (though not perfectly) well.

    It’s also possible that this chart measures CPU production, much the same way, maybe starting at 8,000,000 CPUs for the quarter.

    This is how you make 10% growth look like 40%.

    AMD does things like that a lot in this slide show. Here’s a particular bad example of this:

    yields

    There are two things very wrong with these graphs. Can you name them?

    One problem all of them have is that, again, there are no values placed on the Y axis. You have absolutely no idea what the graphs signify. They appear to be measuring yield (simply because there’s a % indicator in some of them), but with no indicator of values, you don’t know if these graphs are showing a 7%, 70% or 97% yield. Nor do you know if all the graphs are using the same scale.

    That would appear to be the case for the top three graphs. The bottom is missing from all of them. Why, and what does that signify? AMD leaves you clueless.

    Here’s another, this one is supposed to show how much better 90nm is supposed to be than 130nm:

    yields

    This is supposed to tell about the power consumption/frequency curve for 90nm compared to 130nm. Unfortunately, it tells you neither the watts being used, nor the frequencies reached; there’s no values placed on the X or Y axis. This makes it little more than a pretty picture.

    Just in case you’ve understandably come to the conclusion that the guy making the slides at AMD just doesn’t know how to label a graph (and nobody says anything because he’s some exec’s neer-do-well relative), no, they can:

    yields

    They are quite capable of doing a graph correctly (though the next one goes right back to the same old), so rest assured any omissions elsewhere aren’t due to accident or incapability.

    –>

    Conclusions

    These graphs illustrate what I have long observed with AMD. If there’s a choice between being straight or being slick, they’ll choose slick every time. It was true five years ago, it’s true today, if anything, they’ve gotten slicker.

    And frankly, that bothers me.

    I don’t like a company that essentially tries to mislead me and the rest of the world whenever they run into problems. Unlike perhaps many or most, I have a long memory about such things.

    I remember AMD telling an analyst conference that Palominos were going to be out very shortly, then delaying them six months almost immediately after the conference.

    I remember all the “we have no problems making these chips” for months until “oh yes, we did have a problem” situations with the Thunderbird A and just about every flavor of Hammer.

    I remember the “no, we’re not, no we’re not, no we’re . . . , oh yes, we will” when it came to dual-channel desktops, or x86-64 less Semprons or multiple sockets for desktop Hammers.

    Doing things like that may work a couple times maybe a few times more for the less retentive, but eventually people are put on their guard and start looking for the loopholes in anything said.

    Due to what I do, I’ve had to watch them more closely than most, and frankly, I look for the loopholes simply because I’ve found so many over the years. This is simply not a well-run company. Somehow, the product can end up pretty good, but the company itself rarely is.

    CPU making is a tough business, and it’s gotten a lot tougher the last couple years; there’s no denying that. There have been a lot of problems, there’s been some failures.

    What about Intel? They’ve certainly had their occasional issues with honesty in history. They essentially ran both the PIII and PIV designs into the ground despite plenty of warnings of trouble. I still haven’t forgotten the months of denials about MTH on the old 820 boards.

    Nonetheless, eventually they do ‘fess up on major problems and do what they can to make things right. Late may not be good, but it’s still better than never.

    They do try to keep customers informed on major changes to platforms well ahead of time (something AMD may be catching onto when they made good-enough 2006 roadmaps available about a year ahead of time), and they don’t play most of these kinds of games with graphs that you’ve seen here.

    On the whole, I’d say Intel is a better “customer-centric” company, if only to say not-so-good is better than bad.

    Then again, when you have to choose between a first-rate company with a second-rate product and a second-rate company with a first-rate product, it’s never an ideal choice.

    Ed

  • Leave a Reply

    Your email address will not be published. Required fields are marked *

    Discussion
    1. Can you post a link to the article if it's not too much trouble?
      I didn't know that Intel was moving towards processors with less pipelines and less cache, I thought that they would follow the success of the Pentium M's for their future processors which currently have 2 MB cache I believe. However, I will take that back if what I read in the article is different.
      What i meant was that more cache is more helpful on P4 like designs partly because they make lots of failed branch predictions partly because they rely on a northbridge memory controller causing higher latency (so it's nice to have a huge L2/L3 cache). As we have seen on the short pipeline semprons, halving the cache to 256K often has very little impact on performance (hence the 128K semprons). I just thought it was funny that AMD was moving toward designs that would NEED more cache and intel after getting their fingers burnt with the P4 (northwood was ok, but willy and pressie were hardly great) seem to be moving back to a more athlon like design (well more PPro/P2/P3 but certianly not very netburst like).
      I also noticed that DDR2 and PCI-e are only going to have about a 2-year lifespan. I'll stick with DDR1 and AGP for now thank-you.

      I'm pretty sure the PCI-E Gen II thing will still run PCI-E hardware, just you would need a Gen 2 mobo to use any gen 2 hardware. Didn't the original PCI haver versions as well?