Linus Torvalds (torvalds@linux-foundation.org) on 2/6/09 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 2/6/09 wrote:
>>
>>So, I'm curious, what recommendations would you make for
>>improving the Itanium 2 core?
>>
>>the most obvious would be move to a single bundle core,
>>perhaps add OOO. Or you could go niagara style to hide
>>latency...
>
>Realistically? There's not a lot Intel can do that makes
>any kind of economic sense. "Leave the thing behind" has
>been my recommendation since before they even released it
>in the first place.
>
>They can't do the niagara thing, because their only market
>simply isn't remotely interested. They can't do the OoO
>thing (even though I think they should, just to make it
>interesting), since that implies putting a lot of resources
>on it, and their market is too small for that to make sense.
>
>So what Intel appears to be doing is leaving the
>dang thing to die, but releasing these "let's put pieces
>of technology we already had" Frankenstein's monster cpu's
>every once in a while.
>
>I can visualize the meetings: "Let's bolt on QPI, we'll
>just dig it out of the verilog that the Core i7 people
>left behind". Or "Let's tack on another core - that way
>we can claim that it's twice as fast, without actually
>having to design anything".
>
>The most interesting thing IPF ever did was that half-assed
>and sorry half-baked SoEMT thing. And you can still see the
>bolts in the forehead from that one.
>
>So what Intel could try to do is to make a single-banger
>speed-racer design. Put a small crack team of crazy people
>on it, and let them go wild. If for no other reason than to
>at least look alive next to a POWER6. Go for frequency, go
>for simplicity, go for anything: just admit that the
>old core sucked.
>
>Take a chance, for chissake! Intel literally has absolutely
>nothing to lose on it. Go out in a blaze of glory, not this
>disgusting "let's wipe our asses one more time and see what
>we come up with" strategy.
>
>Or just kill the damn thing off. Don't make those ludicrous
>claims about keeping it alive until 22nm. Do anything but
>the "more of the same proven failure".
>
>The definition of insanity is "doing the same thing over
>and over, and thinking that the end result is going to be
>different". And quite frankly, Intel's strategy wrt Itanium
>has been insane according to that simple definition.
>
>Sure, hitting your head against the wall over and over
>might make the wall crumble. But more likely it just
>makes your head hurt. Try another approach, already.
>
>As it is, Itanium just makes Intel look bad. They'd be
>better off using it as a technology demonstration of crazy
>features ("look what we can do in 32nm at 800W in a lab,
>using our experimental Itanium chip") and never sell a
>single one.
>
>But instead, they keep dragging their name through mud by
>comparing themselves with Sparc. It's like Michael Phelps
>crowing about getting a gold medal in the Special Olympics.
>
>Sure, it's a gold medal, but you're just debasing yourself.
>You've shown that you don't have to compete against gimps
>to do well (Core2/Core i7). Why are you competing against
>people who can't do better? Pick on somebody your own size,
>for chissake!
>
>So every time I see Itanium compared against Sparc I just
>feel like Intel should be ashamed of themselves, and that
>the person doing the comparison should just leave it be.
>It doesn't make Intel look any better, and Sun doesn't
>need that kind of "kick them while they're still twitching"
>either.
>
>I'm sorry, I just don't see the point.
>
>Linus

Linus, you're talking about market segments worth billions of $ where x86 isn't present. If the Special Olympics is worth $20B per year and Intel manages to get IPF to a 40-50% share in the next 5 years, was it worth it ? They're currently at around 25%.
If Intel gets 20% of the system price,we're still talking around $1.5-2B per year, out of which 75% if profit since IPF doesn't pay for the FABs.What are the developing costs for IPF ? $400-500m a year ? I doubt they are higher.

IPF is in theory economically sound.It simply needs time to get its foot firmly in the door.
IBM once sold IPF systems.Sales ( systems) were so good, that they were close to HPs ( the irony ) and a threat to Power. Out of the blue ( pardon the pun ) IBM decides to axe IPF systems.Was it because it questioned to whole point of IBM's in house processor effort ?
As the situation of IBM microelectronics gradually worsens it's not unquestionable to think of a future where mainframe CPUs and ASICs are outsourced ( the new Abu Dhabi-AMD foundry, TSMC ) while Power faces the same fate or it is axed. IIRC , there were rumors of Power7 being socket compatible with Opteron.
The same applies for SPARC. Both SUNs and Fujitsu in house processor programs are under pressure.Rock is make or break for SUN while Fujitsu will probably converge Primepower with Primequest in the future, porting Solaris to IPF.