[Beowulf] Three notes from ISC 2006

[Vincent Diepeveen]

----- Original Message ----- 
From: "Mark Hahn" <hahn at physics.mcmaster.ca>
To: "Vincent Diepeveen" <diep at xs4all.nl>
Cc: <beowulf at beowulf.org>
Sent: Thursday, June 29, 2006 2:44 AM
Subject: Re: [Beowulf] Three notes from ISC 2006


>> a huge L3 cache (which most specfp software is somehow) that it destroyed
>
> spec is, after all, CPU2000, and things have changed quite a lot in 6+ 
> years.
> Itaniums were one of the earliest "breaks" of spec cpu2000 codes,
> (besides sun's spec-special compiler).  if you take it2 results
> and remove the ~2 smallest-memory components, you get less impressed.
> I think this has a lot to do with it2's trajectory.

> it'll be interesting to see how much of Core2's performance is due to
> the big 4M shared onchip L2 - Intel doesn't seem to have any 2M scores
> on spec.org yet.

That's why i dislike spec2000 so much in general. Just crafty sometimes 
gives
a reasonable impression from specint. Even then also that program they 
lobotomized to a
hashtable of 2 MB or so for spec2000 instead of like 400MB or so (which is
what it normally runs with at my pc).

I learned that in spec2000 the size of the L2 matters just too much, whereas 
in real life
it hardly matters.

We totally agree here.

>> So getting a 20% higher IPC there than k8 is quite *impressive*.

> hmm, the K8 has an ISA from ~2002, so in the years since then, why
> would you expect no improvement from their competitor?  branch prediction
> _is_ one of the relatively few places where architectural progress is
> being made, at least now that we're over the clock-is-everything era.

Sure, but if you release a new generation car sooner than your competitor 
then don't
find it weird if they buy your new generation car instead of your 
competitor.

>> It's very fair to compare woodcrest to k8, because the next generation 
>> chip
>> from AMD is K8L and as
>> that must use 0.065 technology which will under normal circumstances take

> uh, Core2 is 64nm as well.  you don't really think Intel would try
> mass-producing 90nm chips with 4M cache, do you?

Intel always has a lead on AMD in terms of applying new technology sooner 
and
manages somehow to clock chips higher than AMD.

That has been the case 20 years ago and it still is now.

Additionally AMD fits 2MB cache in the same space like intel fits 4MB in.

Not sure whether that's with woodcrest the case too. It used to be with P4 
the case.

Note that doesn't need to be an advantage for intel. It also means that 
getting rid of heat is more difficult.

>> till 2008 or so to get sold in shops,
>
> that's too pessimistic - it looks to me like AMD's excecutions is pretty
> decent these days, and I expect to see some preliminary K8L results
> this year, and wide availability in 1H07.

Scroll down, in may that roadmap has been changed and K8L has been put to 
2H07 now on paper.

>> Basically first that K8L chip must tape out then it takes another year to
>> produce it and get it in the shops. That's how it normally works.
>>
>> But the combination of new process technology + moving from 3 to 4
>> instructions a cycle will of course give massive problems
>> and headaches to AMD. Especially knowing the years of delay it took to
>> introduce previous technology (0.09) when it was new.
>
> well, AMD's been vehement clear that your pessimism is untrue.

Perhaps you're still assuming some press releases from a few months ago. 
Like april 2006.
Not too long ago, about 1 month, AMD moved K8L to H2 of 2007 at their
roadmap. Shown for example here:

http://pc.watch.impress.co.jp/docs/2006/0503/kaigai267_04l.gif

Would be pretty amazed if newer roadmaps in July 2006 move it backwards to 
H1 2007.

Right now it's H2 2007. Means they get sold by 2008.

>> In short AMD will have to release some quad core k8 end of this year to 
>> be
>> able to compete with woodcrest AND clock it to 3Ghz.
>
> nah.  quad-core requires 65nm as well for decent yields, and it's not like
> AMD's been sitting on their hands for the past 5 years.  I'm guessing 
> they're
> just waiting for the Core2 publicity to subside, then start showing off
> the K8L.

Usually +1 year for AMD. then add some week or 3 and they can sell it.

December 2007, Januari 2008 it's very hard to get the thing sold
in shops before that date.

If they manage i really would be positively amazed!

>> Of course putting 2 more cores to k8 is simpler for AMD than to design a 
>> new
>> core that executes at 4 instructions a cycle.
>
> that's a strange comparison - you must know that K8 is nominally 3x wide,
> as the P6/P-M was, and that Core is basically "P6-L" (+1 decoder, etc)?
> so AMD needs to go from 3 to 4.  tune up the branch prediction, bigger,
> smarter caches, 128b/cycle sse, etc.

Moving from 3 to 4 is not exactly easy knowing that intel already has hit 
3Ghz with their new toy.

Deeper pipelining, influences branch prediction, and not to forget AMD's L1 
is bigger than intel.

Personally i prefer a bigger L1. The 64KB of woodcrest i find quite tiny.

>> That dual opteron dual core 2.4Ghz here is already nearly uncoolable.
> don't be silly.

Have one in your room?

You realize that i'm glad to have an airco here, despite that nearly no one 
else in this country has an airco, in the winter when it was outside -10C
i was running here with airco, just because nowadays there is a dual opteron 
dual core in this room!

For that reason i didn't buy a dual Xeon P4. I just didn't realize that 
nowadays dual core opterons are 95 watt already.

Initial rumours about woodcrest were it would be 60 watt. Now i hear 
suddenly the TDP of those sockets is more like 80 watt.

What is it?

Vincent

p.s. i predict many big universities will mess up with all this confusion 
and fuzz about which processor releases when. i'm not sure
whether inside AMD everyone knows when they go release K8L. Probably a few 
project leaders are using the wip for 24 hours a day now,
and after it tapes out they probably are 20 years older :)