Funny thing, Chic. I got a brother who works at Sequent, which IBM bought, presumably somewhat for their IA64 multiprocessor. He's not a good source of gossip, though, last I talked to him he was repeating the "Merced is just proof of platform prototype, McKinley's the real deal" blather that was current from Intel at the time. That line seems to be in hibernation at the moment.
Merced has been in big trouble for years. My own recollection was that it was originally supposed to ship in '97, that from a Microprocessor Report article from '95-'96 or something. I used to ask CS architecture professors about this regularly, they'd shake their heads. A lot of summer interns worked on it, which may be another part of the problem. The only real specifics I ever got was that too much was being expected from the compiler.
I'm sure Intel could write a book about it, but if I had to guess, I'd say that, maybe, VLIW just isn't worth the trouble. Just offhand, I'd say it's particularly not worth the trouble for server-type usage, which tend to be heavily multiprocess-oriented by nature. VLIW is, in theory, good for extracting fine-grain parallelism out of a single task. The benefit for i/o intensive work is, uh, something of a head scratcher.
Who knows, though, maybe it'll be really good. The other big deal is supposed to be 64 bit addressing, but 64 bit processors have been around awhile now too, the (former) DEC Alpha must be 10 years old. The thing that I find irritating about the long running Merced saga is the stifling effect it's had on computer architecture innovation in general. Sort of like Rambus, Intel decides how the world is supposed to go and everybody loses objectivity. Intel can be wrong, too, it's happened more than once.
Cheers, Dan. |
