"May you live in interesting times." These are, indeed, most interesting times. Battle lines are being drawn for a epic clash that will determine the future of computing. Companies are betting billions, and even their survival, on predicting how it will unfold. Get it right and immense riches are yours for the taking. Get it wrong and, well, you don't want to think about what happens when you get it wrong.

We have four major competitors vying for conquest of our 64-bit future. Let's look at each.

We will start with the easy one first- the partnership of Apple and IBM with the Power Mac G5 machines. IBM makes the CPUs; Apple provides the rest. The G5 is a competent 64-bit processor in its own right and a close second in performance to AMD's Opteron and Athlon 64. It can only run Windows software through emulation and this, by itself, is enough to put it out of the running for mainstream adoption. Mac users will love it but it is doomed to be a success in its own small niche.

Intel and partner Hewlett-Packard have spent years and billions of dollars creating the Itanium. Itanium's EPIC (Explicitly Parallel Instruction Computing) architecture is new and proprietary. If Intel pulls it off they will be able to monopolize the CPU business indefinitely-- permitting only competitors of their choosing and setting their own price for licensing rights to the architecture. With a reward like this Intel has every incentive to do whatever it takes to make it succeed. But despite all their efforts, and much to their frustration, the marketplace has stubbornly refused to embrace the Itanium. Intel puts a brave face on it and each time there is a new release in the Itanium family they claim that it will be the one to finally succeed. It never happens and, in my opinion, it never will.

There are two reasons for Itanium's failure. First, EPIC is a poor architecture for modern CPUs. The problem, and this is a killer, is in the size of the instructions. EPIC offloads much of the logic to the code generation phase. The idea is to let the code generator do some of the scheduling and instruction ordering work the CPU would normally do thus making the CPU's job easier and faster. The downside is this preprocessing requires substantially more code space than comparable instructions in other architectures. By the way- this is one of the same problems that doomed VLIW (Very Long Instruction Word) of which EPIC is a derivative. This flaw is sufficient to spoil the whole show. EPIC was conceived long ago when memory speeds and CPU speeds were at a parity. Requiring more memory to get the same job done didn't seem like such a bad idea but today we have CPUs that vastly outrun memory and the penalty for halting the CPU to access memory amounts to a loss of hundreds or even thousands of instructions. Old father time has passed EPIC by though Intel still isn't willing to concede defeat. In an effort to address the memory problem Intel has been increasing the cache size on each successive release. An upcoming version of Itanium will have- get ready for this! -a mind-boggling nine megabytes of cache! It won't be enough to make Itanium a success. I don't want to make it sound as if Itanium is all bad though. Its floating-point performance is astonishingly impressive. That's commendable but not enough to compensate for its flaws.

The second problem dooming Itanium is the same one the G5 suffers from- backward compatibility. Intel has addressed this problem too but the result is unsatisfactory. 32-bit performance, though better than emulation, is still far from leading edge. It's as if a 32-bit processor was tucked away in a dark corner of the 64-bit CPU. The architectures are radically different and there is no smooth upgrade path for customers wanting to gradually migrate from 32 bits to 64.

In the end, Intel and Hewlett-Packard have spent billions of dollars and years of effort on a processor with poor 32-bit performance, uneven 64-bit performance and an expensive price tag. I have tremendous respect for these companies but in my best analysis this processor is doomed to fail.

Next comes AMD. AMD has taken a different approach with their Opteron and Athlon 64 family of CPUs. They stretched the existing Intel 32-bit architecture to 64 bits and christened it AMD64. The great advantage of this approach is it starts with a proven (and popular) architecture, preserves ideal backward compatibility, offers excellent performance, and provides a perfectly smooth transition path to 64 bits. This is, from the standpoint of the customer, The Right Thing To Do. I described in my journal entry on September 1 how even 32-bit apps benefit by getting up to twice as much memory under a 64-bit OS. All is good, right? Well, not quite. AMD made a couple of costly missteps when they got clever and decided to design in a new bus, integrate the memory controller into the CPU and manufacture it on a new process called SOI (Silicon On Insulator.) The potential to reduce memory latencies, improve SMP performance, increase clockspeeds and lower operating temperatures was too sweet to resist but it was too much to bite off at once. Working the bugs out of the new process caused Opteron to be delayed by a year and while most of the technical risks paid off it did not realize the hoped-for clock speed improvements. AMD should have simply added 64-bit support to their existing Athlon CPU and shipped it a year sooner while they worked on the new, risky technologies for future processor generations. That would have been the boring but entirely sensible thing to do. Perhaps in the end it will work out for the best but it cost AMD valuable time. Still, Intel is nowhere to be seen. AMD is shipping today, the AMD64 architecture is sound and its price-to-performance ratio is unbeatable. Congratulations, AMD.

There is yet one more competitor to examine and it is.....drum roll please......Intel!

Yes, I mentioned Intel already but bear with me. Intel isn't stupid. You don't get to be as big and successful as Intel by missing the obvious and failing to have a backup plan if Itanium should fail. I believe their next-generation Pentium processor, code-named Prescott, contains support for 64-bit instructions. I suspect, though I can't be certain (and AMD isn't forthcoming on this question), that Intel has a cross-licensing agreement that gives them rights to the AMD64 instruction set. If this is true Intel could flip a switch and turn on support for AMD64 instantly at a time of their choosing. Prescott will be released on December 3 and will be in volume shipments in the first quarter of 2004 which will put it just in time for Microsoft's release of Windows XP 64-bit edition.

Time for even more rampant speculation as to how the future will play out. AMD will have the 64-bit market effectively to themselves into the first half of next year. They will be able to command high prices although it will be tempered somewhat by initially low supply volumes. Assuming the tech sector recovery continues it should be enough to help AMD reach profitability. Meanwhile, Intel management has given Itanium one last chance to prove itself. If it fails to catch on once Windows XP 64-bit edition ships, or perhaps even by the launch, they will flip the switch to enable 64-bit instructions and declare their presence. Intel is too stubborn to admit failure with Itanium. Instead they will reaffirm their support for it as the processor for scientific workstations and declare they never really intended it for the desktop anyway. It will simply fade away.

The more things change the more they stay the same. By this time next year Intel will still be on top with defacto leadership in 64-bit desktop computers simply due to AMD's inability to ship in volume. AMD, benefitting from a combination of higher ASPs (Average Selling Prices) due to better penetration in 64-bit server and workstation markets, the cost-cutting discipline of Hector Ruiz and improved economic conditions, should return to profitability and enjoy increased stature in the industry for leading the way in 64-bit computing.

There are some wildcards that could have a major effect on the outcome. It would be a colossal blunder if Intel attempts to force AMD out of the market by creating their own incompatible instruction set. The smart thing to do is adopt AMD64 and show up at the party fashionably late but in great volume. Maybe add a few extra instructions to be a superset of AMD64 just to tweak AMD's nose. If they insist on incompatibility it will set them back (Windows won't be ready for it) and will effectively hand CPU leadership over to AMD. One other thing to keep an eye on is AMD's close collaboration with IBM. IBM is a pioneer in SOI technology and has been instrumental in helping AMD solve their process problems. The relationship between these two companies looks even more cozy as AMD is apparently moving their CPU design team to Fishkill, NY to live and work under the same roof as IBM. It will be interesting to see where this relationship leads. It is not inconceivable to me that IBM may one day acquire AMD. Another wildcard with Intel is heat output with the 90nm Prescott. No one understands how to climb the learning curve better than Intel but they are having serious and unexpected problems with heat generation that could hamper them in raising their clock speeds. This could be a critical competitive issue if AMD/IBM can increase speeds faster.

We truly do live in interesting times.