Two weeks ago I wrote that AMD's Sledgehammer chip might have some advantages over Intel's iTanium, but I didn't explain why. In a nutshell, iTanium's design is complicated, it eats a lot of memory, and its price-to-performance ratio is lousy.

By Rob Landley (TMF Oak)
September 1, 2000

I spent a week back in February on an explanation of microprocessor design aimed at a nontechnical audience. This article more or less takes up where those articles left off, and builds on a story I wrote two weeks ago. The original four may be found at the following links:

Microprocessor Design, Rule Maker Portfolio, 2/22/00
Cold Hard Cache, Rule Maker Portfolio, 2/23/00
RISCing the Pentium, Rule Maker Portfolio, 2/24/00
Merced vs. Crusoe, Rule Maker Portfolio, 2/25/00

I'd say the reason Intel's (Nasdaq: INTC) iTanium isn't shipping yet is that its price-to-performance ratio, which is the speed you get for each dollar spent, stinks. Pentium-based systems provide a lot more bang for the buck, and value for the money is what most people are after when buying PCs. If people want raw performance at any cost, they'll probably buy Compaq's (NYSE: CPQ) "Alpha" processors, or something from IBM (NYSE: IBM).

There are many reasons Intel is having trouble with iTanium performance, but the fundamental one is a design flaw. The iTanium "EPIC" approach to processor design is a greedy pig for memory bandwidth. What goes on inside the processor hasn't been a major limiting factor in PC performance for years; processors are way faster than the rest of the system they plug in to. With on-chip cache and clock multiplying, modern processors now run at about five times the rate of the motherboard they connect to. The real limiting factor is how fast the memory chips on that motherboard can feed instructions into (and accept results from) the processor's on-chip cache. (This is explained in detail in the articles I just linked to.)

Intel has compensated somewhat for the flaw by running extra wires to the motherboard, and by giving systems the option of using faster, but more expensive, memory from Rambus (Nasdaq: RMBS). But, these tricks increase the cost of the system considerably, which leads to the main problem Intel faces: the price-to-performance ratio just isn't there for iTanium.

The remaining problems stem from the fact that, although one of iTanium's design goals was simplicity, it is in fact the first version of a ground-up redesign of all the circuitry, with a brand new instruction set and everything. Even if the iTanium EPIC approach is a fundamentally better way to design microprocessors, Intel still has to gain experience with the new way of doing things and work the bugs out. A lot of people are going to wait for iTanium's 64-bit successor, McKinley, on that basis alone.

A year after the design was completed and delivered to manufacturing, it's still not shipping. Volume iTanium production has been pushed back to the fourth quarter. The prototype chips delivered so far run at about the same clock speed as the faster versions of Intel's bargain-basement "Celeron" line, and the amount of on-chip cache memory (128k) is also at Celeron levels: high-end Pentiums have many times that amount. Intel's Pentium III design has been available at Gigahertz (1000 MHz) and faster speeds since March, a rate iTanium isn't expected to reach any time soon. (Eight-hundred MHz prototypes have been spotted at demonstrations, but not in volume production.)

The small cache, slow clock speed, and delayed production are more signs of design trouble. The iTanium design seems to have turned out more complicated than Intel expected. Reducing cache size is something a manufacturer does to borrow more circuitry to use elsewhere in the design, while staying under the "transistor budget" for the chip (which affects how many chips can be produced per silicon wafer). Producing slower chips is another way to increase production yields, a way to make less-than-perfect circuitry function reliably. Intel's partner Hewlett-Packard (NYSE: HWP) may have designed iTanium to be simple and streamlined, but the actual result seems to be something else.

Compare that with AMD's Sledgehammer, which started from its existing Athlon design and added a little extra circuitry to include a few 64-bit registers and the necessary instructions to manipulate them. Although the CISC design ideas behind AMD's Sledgehammer are theoretically more complex than the EPIC design behind Intel's iTanium, the reality is that AMD's design has a lot less new stuff in it. It's just an Athlon with some 64-bit additions glued on top, so the complexity is counteracted by years of experience and familiarity. This applies to programmers as well: More than 90% of Sledgehammer's circuitry and behavior is the same as an Athlon, so programmers don't need extensive retraining and their development tools don't need extensive rewriting as for iTanium.

Add in the fact that AMD's chips still use a socket instead of Intel's more-expensive cartridge design (a patented technique Intel adopted primarily to preclude competitors from using its motherboards), and that Intel has been encouraging the use of more-expensive Rambus memory to make up for its chips' memory-bandwidth-hogging ways, and Intel has a lot of price hurdles to overcome. Intel may have had a lower cost of production for years due to economies of scale (which is still true), but that's no excuse for getting fat or lazy.

So, although Intel is traditionally the better manufacturer, it may have bitten off more than it can chew on the design side with iTanium. Right now, processor demand is at record levels and neither Intel nor AMD can produce chips fast enough to keep up with the orders rolling in. AMD's traditional position of grabbing Intel's table scraps has changed, because there are far more orders than even Intel can eat, so there's plenty to go around. If AMD can take advantage of this opportunity to catch up with Intel in the marketplace with a better design, grow its production capacity, and pay down some of its debt, this might represent a permanent market shift. AMD and Intel may compete on more equal terms in the future.

As I said two weeks ago: It will be interesting to watch.

-- Oak

Note: Zeke Ashton won't be writing this week or next since we've got him busy on another project -- more details to follow. But, he's in favor of investing this month's $500 in Yahoo! (Nasdaq: YHOO). We'll make a final decision next week and let readers know in advance.


Fun & Folly: Fool Survivor
Time is running out on your chance to play Fool Survivor. Choose the victor from among Richard Branson, Anna Kournikova, Meg Whitman, and the others, and you could be famous throughout Fooldom -- and win one of many groovy prizes.