Intel Again
Bear Argument

Email this article Email this page
Format for Printing Format for printing
Dueling Fools

By Rob Landley (TMF Oak)

I like Intel. I really do. I think it's a great company. But, it is in the middle of a painful business model transition and it's just not communicating its plans for the future clearly.

To find growth, Intel has to diversify. It has admitted this. It has been trying to do it. From Andy Grove's talks about the Internet as the new inflection point for Intel, to its initiative to become a Web-hosting company.

But, everywhere it turns to find growth, it also finds entrenched competition. Intel's Web hosting? IBM (NYSE: IBM) has a big head start in partnership with AT&T (NYSE: T) -- followed by every other Internet service provider on the planet. Internet connectivity through DSL? Cisco Systems (Nasdaq: CSCO) is there already -- as are a dozen highly determined smaller players. Flash Memory? AMD is there in partnership with Fujitsu (Nasdaq: FJTSY). Kingston is there. Hyundai is there. Samsung (Nasdaq: SSNHY) is there.

It's not that Intel can't successfully expand into other niches. It's that it can't hope to establish monopoly margins in any of the new areas I've seen. It's not king of the hill when it's on the road.

In the meantime, this growth is funded by its core business: making microprocessors. But Intel is having problems at home: "Intel inside" is no longer a given. I wrote a Rule Maker column about the Internet's impact on the PC, and the resulting shift in emphasis from "running software" to "connecting to the Internet." Despite the slogans in Intel's ads, a faster processor won't suck the Internet through a 56K modem any quicker. Even with a cable modem, a low-end Pentium can network Napster and Netscape at full speed. Extra memory helps this sort of thing more than extra processor speed.

When America Online (NYSE: AOL) teamed with Gateway (NYSE: GTW) to create an "Internet appliance," they picked a chip from Transmeta to run it.

These days, about all you really need top-of-the-line processor horsepower for is games, and again, a good accelerated graphics card is going to make at least as much difference. At any rate, the PC gaming market has been overtaken by consoles like PlayStation 2 (which uses a custom processor from Toshiba called the "Emotion Engine"). The video game market is bigger in dollar terms than Hollywood, which is why Microsoft (Nasdaq: MSFT) is trying to cash in on it with the X-box. ZDNet noticed consoles were a threat to PCs more than a year ago.

But, let's talk about Intel's core market. I spent a previous duel talking about how AMD's processors were taking market share from Intel. Intel is still fighting its partners over their support of AMD; for example, suing VIA for patent infringement because it makes chipsets for AMD. Micron's (NYSE: MU) trouble licensing Intel's P6 bus for its "Samurai" chipset led it to put that project on hold and instead create the "Mamba" chipset for AMD's Athlons, with eight megabytes of L3 cache embedded in the chipset, resulting in 50% lower latency and sustainable 9.6 gigabyte per second of memory bandwidth. (You may drool now.)

And, I spent a duel criticizing the pathetic morass of lawyers that is Rambus (Nasdaq: RMBS), whose current business model actually seems to be to sue the entire memory industry to extract royalties from a technology they didn't invent and don't even like. Literally. Intel has finally admitted that its relationship with Rambus was a mistake, but it is still contractually obligated to make the Pentium 4 support it. Needless to say, Intel's engineers already said this (loudly and at length), and were shouted down by management. I won't say there was a little bad blood over this, but I will say the only Intel design engineer I knew at the time now works for VA Linux (Nasdaq: LNUX).

So, let's talk products. Intel has milked the P6 core for all it's worth, but when it tried to push it to 1.13 gigahertz it wound up with an embarrassing product recall. Intel's recent shrink to 0.13 micron copper manufacturing should help this, but AMD's Athlons already do 1.2 gigahertz at the larger 0.18 micron size. (Admittedly AMD did go to copper a year earlier than Intel, but how exactly is that a good thing for Chipzilla?)

So, Intel did the Pentium 4. I don't want to be too hard on the P4; most of the really new processor designs Intel introduces need a few versions to get up to speed. (From the Pentium 60 with the floating point bug that turned into a marketing disaster, and the operational temperatures you could fry eggs on, to the wildly successful Pentium Pro that just didn't like 16-bit code.) It's been through it before, it'll survive, but it's generally not fun.

The reason Intel did the Pentium 4 was a technique software people like me call "Loop unrolling." One of the big limiting factors in a chip's clock speed is the ability to synchronize the signals traveling around the chip. The farther they go, the greater the margin of error for exactly when they arrive at their destination. A picosecond here, a picosecond there, it all adds up. The more complex the chip, the greater this uncertainty is, and the more of a delay you need to make it all work right.

The solution is "pipelining," or breaking each instruction down into tiny simple stages that can be arranged linearly along a wire. Figuring out how fast electricity flows along a straight wire is pretty simple, and the small circuits in each stage do as little as possible, so they can be clocked darn fast because there's very little confusion about when signals should be where.

The down side comes in when you have to reset the pipeline. With a 20-stage pipeline, a single instruction takes 20 clock cycles to complete, which is fine when you can have 20 instructions going down the pipe one after the other.

But, if you ever have to flush the pipe (cache miss, incorrect branch prediction -- the reasons are technical, but it happens), you have to restart the missed instruction that many more stages back along the pipe, and flush instructions in between that aren't valid anymore because they'd be executed in the wrong order.

So, the Pentium 4 can be clocked way higher than a Pentium III or an Athlon, sure; but tests indicate that, in the real world, an Athlon or Pentium III outperforms it by around 25% at the same clock speed when running real software. The Athlon's clock speed looks to match the Pentium 4 in the future. Not good for P4. Read more about that at Tom's Hardware Guide.

As for iTanium, go read the AMD duel. It's developed the nickname "iTanic," as in sinking without a trace.

In a day when Pentium III's inability to break the gigahertz barrier is reason to move on to new designs, iTanium is aiming for 800 MHz, and the systems Intel demonstrates at conventions are running at 500 MHz. I know it has 64-bit registers, but honestly, why bother?

On the bright side, rumor has it that Intel's second stab at a 64-bit chip, McKinley, has already taped out (which means the design was completed and sent to manufacturing, so they can start producing prototypes). So, Intel may actually have a 64-bit chip to sell by the time AMD does... assuming AMD doesn't deliver early. But, I still believe AMD has a better design (as explained in Rule Maker columns here and here).

Again, Intel isn't going to be driven out of the processor business anytime soon, but it faces real competition there, something it didn't have before last year. It's no longer a benevolent dictator. It's one among many, and diversifying won't change that. Intel's growth will be as much dictated by what its competitors do (in its core business and in the new businesses it's expanding into) as much as what it does. I think Intel can reach a new balance, but it's going to be a hard road getting there.

Bull Rebuttal »