fluffy grue voted 1 on this story.
It'd be nice to see an in-depth technical piece comparing two pieces of hardware, but this is still somewhat lacking. For the part where it is actually making a comparison, it's quite unobjective. For starters, IA-64 is not VLIW, but a sort of RISC-CISC-VLIW fusion called EPIC. It still has a scheduler, it's just that a lot of the burden of scheduling algorithms are largely offloaded onto the compiler. There's still a lot of run-time scheduling going on, though there's no longer any out-of-order execution.
In addition, their predication vs. speculation example is a little biased. The point to predication isn't the ability to combine comparisons into as few operations as possible, but to remove jumps for small conditional blocks. With speculation, the Alpha still needs to perform a costly jump, whereas with IA-64's predication, the instructions within the conditional block are simply ignored if the condition is unmet. When you've got a deep pipeline, basically converting instructions into NOPs is a hell of a lot more efficient than having to jump. Compaq's example certainly doesn't reflect that.
Their benchmark numbers are definitely biased, given that they don't even SHOW IA-64's performance. Rather Orwellian, if you ask me. Also, comparing roadmaps isn't exactly a fair comparison of underlying technologies.
Basically, what VLIW and EPIC do is offload the expensive, convoluted and highly-bug-prone scheduler circuitry into software. Underneath every modern RISC and CISC CPU is basically a VLIW core. The AMD K7 is a perfect example of this; all that its scheduler does is convert CISC into very fine-grained VLIW. The Transmeta Crusoe is an even better example: the scheduler is completely implemented in software. Technologically, the core of the Crusoe isn't really that different from the post-scheduler pre-commit core of the K7, PowerPC, P6, or Alpha.
In any case, the extent to which this article is a comparison between the Alpha and IA-64 is very minimal; basically they make some false claims about IA-64, show their lack of knowledge in what the IA-64's predication actually does, and then go on to make random musings about their architecture which are irrelevant to the "debate."
Don't get me wrong, I like Alphas. I'm also not exactly head-over-heels with the IA-64 architecture. My favorite emerging CPU architecture right now is the Crusoe, but not because of the company or people behind it but because they've finally figured out that hardware-based instruction scheduling is limited and needlessly costly, not to mention that having a 24GHz CPU is pointless when there's so much more useful and interesting stuff to be done with a "lowly" P2-500-class CPU in embedded applications. However, I won't stand by and agree with arguments which aren't based in reality.
"Is not a quine" is not a quine.
I have a master's degree in science!
[ Hug Your Trikuare ]