As for whether it's a replacement for the Intel/AMD chips, that's a
definite negatory. The CPU market is one thing, but the x86 CPU market
of which Intel and AMD are the two biggest members are quite another
market. No matter how much performance a competing architecture brings
to bear, it's still never good enough to be considered a replacement
for x86. The only thing that's good enough to replace an x86 is
another x86.
that's a bit of an overstatement - transmeta has found a corner of
the x86 market where a non-x86 can do some damage. sure, they haven't
taken over the world, but they're certainly viable.
Just look at the relative success of Opteron vs. Itanium.
it's very intresting to compare. for instance, the it2 has been
pushed very hard ahead of demand, and hasn't exactly shone.
but the opteron has been anti-marketed by AMD (ie, yes, it does
run windows, but most windows-on-opteron is in brain-damaged
Intel-compatible mode). so the Opteron has a big bump to performance
just waiting for someone in Redmond to take advantage of it.
the it2 desperately needs a perf bump, since at least one large
company has staked its enterprise line on it. and the it2 perf bump
doesn't seem forthcoming.
look at the routes these chips took. opteron tweaked the existing,
fairly boring athlon core, put a nicer cache on it, great memory interface,
and elegant system infrastructure. considering how little AMD has improved
their core over the years, their success is amazing. quite a contrast
to the it2, whose raison-d'etre was to do a new core ISA from scratch.
Intel/HP have always claimed that EPIC's design was motivated by wanting
to scale performance into the stratosphere. alas, for current it2's at
least, where performance is good happens to be places where the entire
application+dataset fit into cache. would it2-9m be faster than an
opteron with 9M cache? I doubt it. what does that say about Intel/HP's
innovation-in-ISA?
and multicore is not going to save it2. it'll give a nice tweak to
performance under some loads, no doubt. but since it2's competitors
will probably go multicore before it, Intel gets no relative advantage.
and after all, it2-multicore is essentially just making up for the fact
that the it2 ISA designers are so hostile to out-of-order. one of those
mysteries - if OOO is such a bad thing, why have other chip design teams
been able to push it so far? no doubt the OOO designs have been harder
because of the OOO, but has the it2 taken advantage of this to push other
aspects of performance, or to have a shorter time-to-market? unless you
count obscene onchip caches as an ISA innovation, I just don't see where
the it2 is winning.
here's a match I'd love to see: AMD buys Transmeta. why? mainly because
Transmeta has higher-order, firmware-based insight into performance.
for instance, Transmeta can implement using a hard+soft combination,
more intelligence in predicting branches, or in predicting computed ld/st
addresses, or even in predicting values. suppose you can accurately predict
when a store writes to an address which won't be used any time soon. why
not write it through to dram? no sense in wasting a cacheline on that.
predicting loads, of course, is even better, since they tend to stall,
but stores can be posted. but can you afford to treat every load the same
way you do branches (with a multi-state FSM to predict the outcome)?
perhaps not in hardware, but maybe in firmware (maybe even encoded in your
internal instruction stream, since AMD already does cache partially decoded
instructions, though doesnt' go as far as Intel's trace cache...)
regards, mark hahn.