slacker (s@lack.er) on 2/6/09 wrote:
---------------------------
>Michael S (already5chosen@yahoo.com) on 2/6/09 wrote:
>---------------------------
>>Never ever in Itanium history there was 20% lifetime increase. Actually, I don't
>>remember any increases at all although it's possible that I forgot about some 1% case.
>>On the other hand, the 10% lifetime decrease (McKinley 1000->900) is well documented.
>
>I think this was its history:
>
>Merced - 800MHz
>McKinley - 1000MHz -> 900MHz (buggy)
>Madison - 1.5GHz
>Madison 9MB L3 - 1.6GHz -> 1.66GHz

AFAIR, 1.66GHz version of Madison-9M was available from the beginning, but only at 333MHz bus (more commonly knows as 666 MHz). At time of launch no vendor had 333MHz infrastructure in place and even later on Toshiba was the only one. That's the reason why 1.66 GHz benchmark results appeared later than 1.6 GHz.

>Montecito - 1.6GHz -> 1.66GHz
I am pretty sure that Montecito never exceeded 1.6 Ghz.

You forgot another line:
Montvale - 1.66GHz -> 1.66GHz




>
>I have three reasons for having faith in Intel this time around. First, they've
>completely given-up on the power budget issue, raising TDP all the way to 170W.

Which is not much taking into account four cores and all new uncore stuff. Remember that Montvale is 104 W.
In fact. with proposed less aggressive low-level circuited technique 2GHz at 170W already looks like very aggressive target. Unless they finally made Foxton replacement working.

>This should give them the power legroom required to raise frequency.
>Second, they're
>re-using the same core from Montecito (porting it to 65nm), so there's no additional
>core complexity to bite them in the rear.

But also means that 65nm opportunities are not fully exploited.

>Third, the new system interface is 100%
>static, synthesized logic, and they claim it was over-engineered well beyond a 2.4GHz target.
>

The core is the problem, not the interface. If they want a core running at high frequency they have to give up on one of the four:
1. Core count.
2. Wide issue.
3. Full forwarding network and ortogonal register read ports (i.e. revert to design resambling a true VLIW instead of current RISC in VLIW clothes).
4. Short pipeline
Since they're doing neither (and, IMHO, rightfully so) they should forget about hi frequency.


>I'm using superficial reasoning here, but I think Intel and HP will be in a mess
>if Tukwila never exceeds 2.0GHz. I have no faith that Intel will actually release
>a 32nm Itanium within 6 months of 32nm x86 processors, and probably not even within 12 months.

Did they ever said that it would happen in 12 months? Did they ever said that it would happen ahead of 23 nm x86?
I don't remember anything like that said by Intel officials. They said that they are skipping 45 nm which are bad news especially taking into account that Intel 45 nm process has very real advantage over 65nm. The rest are attempts to rationalize bad news into good news.

Despite that I don't agree that lifetime 20% frequency increase matters and worth seeking for.