Saturday, March 11, 2006

IDF summary report

NGMA is Pentium 3

I stated repeatedly that Intel NGMA is Bob Colwell's Pentium 3. A lot of Intel fans strongly opposed such a notion and insisted that NGMA would be all new from ground up.

I was 100% positive, because I know this: it is impossible for the Israeli amateurs to do some radical new stuff out of no where, they simply don't have the track record for ground breaking work.

Now it's all confirmed, the Core architecture (NGMA, Merom/Conroe/Woodcrest) is 95.5% Bob Colwell Pentium 3, plus 4% Bob Colwell list of todos when he left Intel, and probably 0.5% Israeli patchwork.

Look at those minute improvements: smarter cache, larger cache, better chipset, faster bus, grouping a couple of adjacent instructions when doable, Intel deeper sleep, tweaking a few SSE instructions, issuing 4 instructions per clock....I am trying to put all I remember here, but none of them is mildly interesting stuff. AMD would not even mention such minor improvements with their core revisions.

One thing that shocked me was that Anand reported that with the NGMA core, all SSE instructions are executed in one clock cycle instead of 14 clock cycles. I thought how could that be possible? Did Intel Israel invent a special 1 stage SSE unit? It turned out the cheerleader was too eager to cheer. It was all hoax.

It's bad that Intel Israel bragged so much for so little. And it's not good that they willfully failed to mention Bob Colwell the American, but instead took every credit of the Pentium 3 Israeli edition as all new.

Conroe Performance is a hoax

There is simply no way a 2.66GHZ Conroe can be 20% faster than a 2.8GHZ AMD64. Woodcrest is 10-20% slower than Opteron 280. INTEL had to use a 3GHZ Woodcrest to pit against a 2.4GHZ Opteron 280. Woodcrest is pretty much identical to Conroe, both are based on Merom.

NGMA/Core/Merom/Conroe/Woodcrest is 32 bit

One thing I observed was that there was no demonstration or even indication of AMD64 compability of the NGMA chips during IDF. No Windows Vista demos, no Linux 64 bit demos, no CPU-Z screen capture indicating the presence of x86_64. In HP's demo of the DL 380 with two Woodcrest processors, there was only 2GB of memory, odd for a server of this calibre.

If the Cores were 64 bit capable, it would be unthinkable for Intel to not demonstrate this capability, which is critical for Windows Vista and enterprise computing. The fact there was zero sign of AMD64 capability means all NGMA chips as of today are 32 bit only.

While AMD's grand masters grew up with 64 bit, the Israeli amateurs don't even understand there is a need for it.

The dire situation for Intel

I predict that Merom/Conroe/Woodcrest will be delayed. Intel will struggle to get AMD64 compatibility figured out. One way is resort to emulation, but there will be 10-20% performance degradation, as we can see from current EM64T benchmarks. AMD64 runs 10-40% faster in 64 bit mode.

There will be no IMC, no CSI, no true multi-core for INTEL in the forseeable future.

On desktop and mobile, AMD's preemptive strikes in May and June will be many steps and many months ahead of Intel's.

On servers, Woodcrest/Bensley will be stuck at 2P level for a long long time. In the end, Intel may pay big money to IBM to get the Hurricane chipset so they can get 4P working, but the cost will be very high. On cost alone, there is no way for Intel to compete against AMD64, which can do glueless ccNUMA up to 16P.

The biggest problem for Intel is, time has already run out. With FAB36 converted into 65nm, AMD will be able to supply 50% of the global PC market by 1Q07.

Intel is and will be five generations behind AMD64, IDF reconfirmed that.

Game over!

14 Comments:

Anonymous Anonymous said...

"There is simply no way a 2.66GHZ Conroe can be 20% faster than a 2.8GHZ AMD64. Woodcrest is 10-20% slower than Opteron 280. HP had to use a 3GHZ Woodcrest to pit against a 2.4GHZ Opteron 280. Woodcrest is pretty much identical to Conroe, both are based on Merom."

Sharkboy simply can't take it that the Conroe is and will continue to be faster than AMD64. He just can't take it.

9:11 PM, March 11, 2006  
Anonymous Anonymous said...

i've been reading your post ever since the release of conroe's benchmarks, and i went back and read most of your posts. honestly, i'm really impressed by how much work you've done to support your points. i want to ask, if there is any info out there that explains merom's architecture?

9:20 PM, March 11, 2006  
Blogger netrama said...

Conroe and what ever crap Intel will sell , is six months away ...some the websites even had a few pics of
a "Quad Core" Hyper Processor (hype processor) called "Kentsfield".
What I am curious is..if the Intel boys have already made the engineering samples of all these hype devices.. what is their engineering teams working on right now ?? Two things they atleast look skilled in so far is : 1.) Cut and Paste 2.) Tweak the instructions to fool the benchmaks..

10:02 PM, March 11, 2006  
Anonymous Anonymous said...

i agree..
my friend who works in OEM industry in Taiwan (they supply for Dell) just told me that merom's progress is really slow. she said that they've just solved the heat problem, and the performance is really poor (probably as poor as the early stage in yonah 1's pre-production. back then, yonah has to work on large 12inch fan, spinning @ 5000RPM the least). i really wonder how intel pulled off that benchmark.

7:49 AM, March 12, 2006  
Anonymous Anonymous said...

Dell is selling Xeon server for GBP 299, and laptop for GBP 299 in London. Dell must be very desperate to get rid of its inventory. Trouble sign.

Regarding to to Intel NGMA, that is not really a new design for the future. Without a proper bus architecture, it is more like a temporary fix for a couple of year at max. Intel could never catch up AMD in 2 years time.

9:56 AM, March 12, 2006  
Anonymous Anonymous said...

How about this Merom benchmark...Computer Base.de

I belive that Merom/Conroe IS 20% faster than Yonah which is equal to X2 performance.
AMD must deliver 30% more performance per clock or just more Mhz with the same TDP.
Cache, adjusted pipeline, 65nm - doesn't matter.
But this is a name of the game...

11:47 AM, March 12, 2006  
Anonymous Anonymous said...

Your rant is laughable.

Here is Merom running Windows x64 edition:

http://tweakers.net/ext/i.dsp/1141761897.jpg

12:42 PM, March 12, 2006  
Anonymous Anonymous said...

could somebody confirm whether Intel's new chips
are single thread per core, or more.

1:42 PM, March 12, 2006  
Anonymous Anonymous said...

Conroe is 64 Bit, as is Merom, as is Woodcrest. Conroe uses half the energy of the best AMD parts and it does indeed benchmark better than 20% in almost all cases.

Where is AM2? Wasn't that due out in January? or wait February? no... hmm are they saying July now? Wait AM2 actually decreases performance at the same clockspeed? Wow yeah AMD is screwed for the next 8 months or so.

65nm is going to improve the performance of AMD chips right? When was that happening again? oh yeah 2007 right? Intel has been shipping 65nm with the best yields they have ever seen since October 2005 and now they have a new incredibly efficient design (Core Microarchitecture) that they will bring to bear with their 65 process and beyond iwth 45nm which appears on track for 2007/2008 products.

I'll happily say that it is indisutable that AMD has had better benchmarking ships for the past year or so. But your whole "analysis" blog is blind AMD fanboy devotion.

Competition is good for the consumer, and Intel has just (finally) stepped up to the plate.

5:00 PM, March 12, 2006  
Anonymous Anonymous said...

hey sharikou, you saw this?

http://torrentfreak.com/the-piratebay-wins-6000-thanks-to-tv-show-contestant/

even the guys running piratebay are using opterons and HP is selling it to them!

6:17 PM, March 12, 2006  
Blogger Sharikou, Ph. D. said...

Joshua,

Thanks for your insight.

9:10 PM, March 12, 2006  
Blogger Sharikou, Ph. D. said...

Note:

Yonah or Merom's shared L2 cache simplied cache coherence problem for a very special situation, which applies to 2 cores on die only.

However, when you use two Woodcrest dual core CPUs to form a 2P SMP, the benefit of simplified cache coherence will be suddenly gone, because even though the two cores inside one CPU are sharing the L2, they must still go out to the FSB to make sure that the other CPU is coherent. That's why I said Intel's NGMA is simply a special hack that can't be generalized, it's another dead end.

10:20 PM, March 22, 2006  
Anonymous Anonymous said...

Wow. I have to say, where I was skeptical about your claims before, every point you have made re: conroe/NGMA from Intel seems to be on the mark... The evidence is all there if anyone truely cares to dig enough and really see all the evidence. Nice job!

Also, does anyone know if the HP 380/Sun Opteron comparison at IDF included any benchmarks?

2:10 PM, April 12, 2006  
Anonymous Anonymous said...

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2748

11:07 AM, May 01, 2006  

Post a Comment

<< Home