Advanced Micro Devices
What Does the Future Hold?

Format for Printing

Format for printing

Request Reprints


By eachus
June 14, 2006

Posts selected for this feature rarely stand alone. They are usually a part of an ongoing thread, and are out of context when presented here. The material should be read in that light. How are these posts selected? Click here to find out and nominate a post yourself!

I've stated many times here that I am by nature a futurist. Where technology is concerned, I live on the bleeding edge, and try to predict where the bleeding edge will be in a few years. Of course, when investing I can only use that knowledge to tell whether current pricing of a technology stock is based on factors that won't hold forever, or better, that an industry (not a particular player) will never recover, or will do spectacularly well.

Intel vs. AMD is currently an exception to the rule. Why? Because right now AMD and Intel are planning for different futures for the whole computing industry. So you get to place your bets on different futures. It is of course, worth pointing out that I haven't changed my current position on both stocks--out. Less than six months ago, I said that AMD was more likely to sell both above $40 and below $20 this year than to do neither. I'm still not sure whether AMD will get that low, but I still expect to buy back in below $25. Intel is probably closer to their 52 week low, and when the latest round of piling on by analysts finishes, Intel should be a good medium term buy.

But where will they be in three years? Or thirty? That may be a much different story. It is worth establishing the current and near-term positions of both companies, then look at where their strategies are carrying them.

First Intel. Technically speaking Intel is near bottom. Yes, we all know that Conroe, Woodcrest and Merom are coming. However while this new generation puts Intel a step ahead of AMD, it graphically illustrates how bad their current offerings are. So that is the place to start. I don't know how much of a life the Celeron brand has left, but it is much shorter than the Pentium brand, and both will be gone by the end of next year. Not only from Intel's price lists, but from advertisements for low-end machines made up of leftover parts. They may have a Pentium 4, Pentium D, or Celeron CPU--but they won't advertise it.

Why not just drop Netburst production altogether? The problem is yields and binning. I don't know if Intel could get high enough yields on the 2 Meg L2, low clock speed version of Conroe this year to flood the market with it--but Intel's plans seem to say no. I am sure that will happen next year. I also expect that Intel will be able to get their yields and binning on the high-end Conroe parts (including the extreme edition) to push AMD out of the catbird seat, if K8L is late or doesn't measure up.

In mobile, we are beginning to see what was once unthinkable. Manufacturers are switching to AMD Turion instead of Core (Yonah). Why? It seems to be a combination of factors. Yonah is not 64-bit capable, and will have a short market lifespan. Merom, which is 64-bit capable is going to require a new chipset. So OEMs and ODMs are introducing a few Yonah models, and filing out the line with Turion, Athlon64-M, and mobile Sempron models for now. Will Merom end this window? Probably not. Merom will be close in timing to K8L mobile chips, so that is probably a wash, but Merom will start at the top and migrate downward. It may be two years or more before Intel has a low-end 64-bit laptop chip. So my read is AMD will get a lot of the low-end laptop business, and Intel will hold the (large) business and high-end segments.

What about servers? It amazes me that Intel has been leaking Woodcrest news and benchmarks so liberally that they basically killed Tulsa and Sossamon before they were announced. The only reason it didn't kill Dempsey is that Woodcrest is promised to be a drop-in replacement. Is Woodcrest really that good and worth waiting for? Yes, especially compared to Intel's alternatives. A dual-socket Woodcrest system will outperform a four-socket Tulsa system, and low-power versions of Woodcrest will definitely outperform Sossamon, in addition to being 64-bit.

There is one other thing to keep in mind about Woodcrest. Intel's starting lineup ranges from 1.6 GHz up to 3.0 GHz, almost 2 to 1 in clock speeds. In a few months, the only question when buying an Intel multi-socket server will be which Woodcrest clockspeed to choose. It will take a few quarters for this to work its way through the various Intel server vendors, but once that happens, Intel will be gunning for dual-socket Opteron business.

On to AMD. For this year, AMD has announced their new (rev F) desktop lineup (yawn), and the DDR2 mobile lineup as well. (Not yawn, as explained above. Intel has left a window open in the low-end mobile space.) The rev F servers have yet to be announced. Currently no surprises are expected. Late this year, AMD is promising to start shipping rev G desktop chips, and again no surprises are expected. Perhaps the kindest way to describe AMD's new lineup is setting the stage for next year. AMD needs K8L (or better) to challenge Intel's coming lineup at the high-end. Even if rev G does have some IPC goodies (load-store reordering, longer decode buffer, faster caches, 128-bit single clock SSE support, etc.), rev G won't be announced this year, or there may be a few specific vendors with high-end systems at Christmas. Rev G, and rev H will be next year's story.

So there will be switches in performance leads, an opportunity for AMD to pick up low-end business, and in a year or two everything will be back to parity. Right? Wrong. That's what makes being a futurist so much fun. ;-) First, Intel has basically abandoned the high-end x86/x64 server market. (Poor Tulsa, stabbed to death in the womb.) AMD on the other hand is planning to make 'glueless' 64-core Opteron systems possible, and 32-core systems very reasonable. Those products are a year away, at the least, but AMD is continuing to move Opteron into the enterprise server space. That is not news. What is new this month, and requires putting the pieces together, is AMD's push towards attached (FGPA or otherwise) processors.

For example, take the 4x4 announcement. You and I know that right now, there is no advantage to more than two cores for gaming, so what is going on? Take a look at the pictures: There is something missing. There is no HT link from the second CPU socket to the chipset. Why? My guess is that the 4x4 announcement and motherboards are really part of the Torrenza program. Certainly if I bought a 4x4 motherboard, I'd be looking for an FPGA or co-processor to stick in that second socket. Why not an HTX slot instead? Think choice. Some co-processors are not going to need large amounts of memory bandwidth. Those will go in HTX cards. Others, especially the kind of linear algebra accelerator that interests me, will require plenty of memory access. The dual-DDR2 memory channels that go with an AM2 socket will certainly fill that need.

For bleeding edge people, this era has arrived. When will it become just leading edge? Any day now. Those willing to be on the bleeding edge have been using FPGAs with current generation (DDR/Socket 940) Opterons. I won't be surprised if the delay in the rev F/DDR2 Opterons is to bring a couple of co-processor vendors on board at the launch. However, from a stock owner's perspective, bleeding and leading edge is uninteresting, a significant market presence is necessary--or should be necessary :-( to affect the stock price. That will probably occur next year. (However, expect lots of leading edge (read HPC) sales between now and then. Sun and Cray have already sold bleeding edge versions, and IMNSHO when AMD announces an ACML version that supports at least one FPGA accelerator, the technology should be considered off the bleeding edge and into practical use.)

Where does this leave Intel say three years from now? No clue. I see three possibilities, but I don't think any is likely. One is for Intel to abandon the high-end entirely. Unlikely, but remember poor Tulsa. Another is for Intel to license Hypertransport, including coherent Hypertransport and use it across their CPU line. Even less likely? I think so. The remaining possibility seems to be for Intel to get CSI out the door in jig time, and get either the same coprocessor vendors supporting HT or their own set of coprocessor vendors on board. The problem with that alternative is that, at this point, the only common in CSI is IA-32 and IA-64. Widening CSI to support multiple coprocessors will probably take time--and the clock has already been running for years. Intel, or someone, could build a CSI to HT bridge, but that would add latency, and the whole reason for using HT instead of PCIe or Infiniband is to keep the latency low as possible.

Hmmm. That's what I really wanted to say, but of course, there is more. Part of it is the whole question of how much PC capacity do you need. Right now, a decently configured Socket 939 system can support a high-end video card with no problem. In fact, a single-core 3800+ can support an SLI configuration of video cards pretty well. Dual-core is a "nice to have" right now, even for high-end gamers. So over the next few years, there is nothing on the software roadmap (with the possible exception of Windows Vista) that will stress current systems. Cheap systems sure, but part of what is going on is a move from cheap systems to inexpensive but very adequate systems. Low-end Conroe and AM2 systems, assuming decent memory (even decent DDR2-533 memory) and a decent graphics card (nVidia 7300 or ATi 1600) will be very decent gaming machines. They may not run 1600x1200 with all the eye candy on, but then again, they will mostly be sold with 1280x1024 or lower resolution flat-panel displays.

So one trend to watch is the accessibility of decent PCs around the world--along with decent (possibly wireless) Internet access. PCs will become much more ubiquitous. (This is different from ubiquitous computing where everything has computational technology--but we are getting there too.) So I think that AMD and Intel can continue to sell hundreds of millions of CPUs for the indefinite future. However, I don't see how Intel can continue to increase revenues beyond 2007 or 2008 by selling CPUs. Currently the market for CPUs is highly elastic. (A small decrease in price results in a large increase in demand.) But at some point, the other components in a PC mean that lowering CPU prices won't have a large effect on demand. AMD can, by dropping prices, try to take a larger share of the CPU market away from Intel, but for Intel that is not a possible strategy--in spite of all the headlines about Intel's recent price cuts. Those have everything to do with the transition to Conroe, and do not constitute a price war. (Intel has to find (relative) prices where their customers will buy the CPUs that Intel is producing. In this case Intel has decided to price high-end Conroe CPUs to sell, and everything flows from that.)

In servers, and where AMD is going with co-processors, the market will continue to expand, almost without limit. Again, speaking as a futurist, have you recently bought something where the company who sold it spent zero seconds of employee time processing your order? Probably, I know that I have. In fact, when buying software it is becoming commonplace, not just for the company doing the sale, but for the delivery to be completely automated as well. This makes for great productivity, and that is the engine that makes the economy go round. It all requires servers, and the better the servers get, the better the company will do. Let me give you just one example. Say you are at Newegg's website buying a new video card. If Newegg's computers can figure out what you are likely to buy with the video card, and display that, they make another sale. Or if you ask to see say nVidia 7600 family products, ordering the display to put models other people chose first, or those that are in stock, may make a sale more likely. All this requires not just an automated sales system, but one that is as intelligent as possible. The incremental cost of additional CPU cycles is relatively small, so once better software is available, it is a no-brainer to field it.

Become a Complete Fool
Join the best community on the web! Becoming a full member of the Fool Community is easy, takes just a minute, and is very inexpensive.