DailyTech - Intel Lays Down Path for 32-Core Processors

Submit News

IT Intel Lays Down Path for 32-Core Processors
Tuan Nguyen (Blog) - July 10, 2006 5:18 PM

26 comment(s) - last by Motley.. on Jul 13 at 5:16 PM

Conventional CPU package

New bumpless build-up layer (BBUL) packaging

BBUL sample

Intel may finally use BBUL

Recent reports suggest that Intel is working on research for 32-core CPUs by 2010. The new processor will contain a total of 24MB of last-level LL cache split among processor nodes. The CPU is anticipated to contain eight processor nodes with four cores per node and 3MB of LL cache to each node. Intel says that BBUL packaging method (which was actually announced back in 2001) will be just as crucial to attaining these specifications as is reducing transistor size.

Before reaching 32-core processors, or even half that, a number of technologies must still be developed to address key bottlenecks in manufacturing processors. Specifically, Intel will finally be developing new processor packaging techniques over the next several years using BBUL. The new packaging technique will allow Intel to make a number of achievements including:

Increase core count
Increase processor frequency
Add other core types such as chipset or other dedicated controller
Lower voltage
Reduced inductance and electrical noise
Thinner, lighter packaging
Significant power savings

Intel says that today's packaging requires too many layers, and the processor core is exposed, making it prone to damage. Current processors place the die on top of a package, connected by C4 bumps, just like ball grid array (BGA) chips. The packaging layer itself consists of three separate layers: two outer copper interconnecting layers sandwiching a center plastic core layer with holes drilled by lasers and filled with copper. The pins that actually allow the processor to interface with a motherboard socket are attached to the bottom copper interconnection layer.

With BBUL, Intel removes top interconnect layer altogether. The processor core is actually embedded into the core layer with only its top surface exposed. The bottom interconnect layer is built-up, or "grown," from the processor core. One major limiting factor that BBUL eliminates is the need to use C4 bumps to connect the processor die to the packaging. As processors become more complex and contain more cores, more interconnects are required and as a result, C4 bumps are hitting their limit in terms of size and density.

With regard for the new massive-multi-core CPUs, other details suggest that Intel will be using more optimized "mini" cores. These cores will focus a great deal on parallel thread processing and be void of floating point processing units. Using BBUL technology, Intel will be able to embed high-speed dedicated math processors into the same package.

Some of this technology already exists today care of Sun Microsystems. In a previous report on Sun's T1000 and T2000 servers, Sun introduced the UltraSPARC T1, a multi-core processor supporting up to 8-cores. The UltraSPARC cores focus on processing concurrent threads, executing up to 4 threads simultaneously per core. A math processing unit is in the same physical die, but is not within any single core.

Along with such technologies like extreme-ultraviolet (EUV) lithography, smaller silicon gates, better interconnects, BBUL, High-K gate dielectric and better processor designs, multi-core processors are quickly becoming more mainstream. While 32-core processors are still a few years away, Intel will be introducing Kentsfield, its first quad-core processor early in 2007. Likewise, AMD will also be introducing quad-core Opteron "rev. G" processors in the same time frame.

Comments

Threshold

Username
Password
remember me

This article is over a month old, voting and posting comments is disabled

32 Cores by 2010

By OmegaPrime on 7/10/2006 9:33:55 PM , Rating: 2

That seems like an amazing feat for Intel to try and get 32 cores on one chip by 2010. What I'm wondering is how much power such a chip would use, as well as how you'd cool such a beast. Even with Conroe chips being more efficient than previous Intel iterations, I somewhat doubt air cooling would work well.

RE: 32 Cores by 2010

By Knish on 7/10/2006 9:35:09 PM , Rating: 5

As someone said on Aces.

"Intel also told us we'd have 10GHz processors by 2005"

Just thought I'd add that.

Parent

RE: 32 Cores by 2010

By Questar on 7/10/2006 9:47:35 PM , Rating: 2

Wow! I never heard of that! Damn lying Intel!!

Parent

RE: 32 Cores by 2010

By hstewarth on 7/10/2006 11:33:42 PM , Rating: 2

Things have change in speed ratings and cores. Well can you actually compare a 3Ghz Woodcrest to a 3Ghz Pentium 4. 1.5Ghz Woodcrest is actually closer to 3Ghz p4 ( single core of Woodcrest ). So taking account of that speed a 3Ghz Woodcrest and 2 cores a 3Ghz woodcrest is likely 4x the speed of 3Ghz p4 - so in p4 terms - its could be consider a 12Ghz P4.

One things that likely will change is development of software, since Multi-core is getting really common place and more popular, multi-threaded applications will be come more common place to take advantage of technology.

Also we not sure if the 20Ghz is single core, or total power - with 32nm its could be single core. Its going to happen - but I would guess the date is off - probable a year late. Just my guess.

Parent

RE: 32 Cores by 2010

By Phynaz on 7/11/2006 11:57:31 AM , Rating: 2

I'm sure he was being sarcastic.

Parent

RE: 32 Cores by 2010

By saratoga on 7/11/2006 2:58:10 PM , Rating: 2

quote:
So taking account of that speed a 3Ghz Woodcrest and 2 cores a 3Ghz woodcrest is likely 4x the speed of 3Ghz p4 - so in p4 terms - its could be consider a 12Ghz P4.

I realize what you're trying to say here with this analogy, but its not really correct. Hz is a measure of frequency, not performance. The correct way to phrase this would be to say that its 4 times as fast, because in actually a 3GHz Woodcrest has a clock speed of 3GHz, even in P4 terms.

Also, saying that two cores are twice as fast as 1 is extremely optimistic. Its almost true in certain applications, but not in general.

Parent

RE: 32 Cores by 2010

By Enron on 7/10/2006 10:29:42 PM , Rating: 2

Then they realized that more work per clock is the better approach. 2010 is still far away.

Parent

RE: 32 Cores by 2010

By deeznuts on 7/11/2006 12:50:06 AM , Rating: 3

And everyone accepted that as accurate? I remember reading about the 10GHz claim from Intel, but never that it was supposed to happen in 2005. Sounds like some spinning to me.

http://www.theregister.co.uk/2000/12/11/intel_plan...
This link at first appears to state Intel claim 10GHz by 2005, but it was written in 2000, and says that Intel claims 10GHz in 5 - 10 years. Not exactly a claim that it will be available by 2005.

http://www.theinquirer.net/default.aspx?article=74...
I found that, but looks like speculation apparently since the language said they have a note which they won't declare how they got it.

Where has Intel declared they would hit (not could hit) 10GHz by 2005?

Parent

RE: 32 Cores by 2010

By Knish on 7/11/2006 12:53:23 AM , Rating: 2

Well if you read carefully no one at Intel has officially claimed this either. Some guy at THG says so though.

Parent

Supporting Links/Clarification

By dilz on 7/11/2006 2:36:33 AM , Rating: 4

Reference to a ZDnet article from 2000 that predicts 10Ghz by 2011. The article's link defaults to ZDnet's home page.
http://www.geek.com/news/geeknews/q22000/chi200072...

The nugget we've all been waiting for. I hope I have your attention...

"Grove Predicts 10-GHZ Chips"
Brooke Crothers and Jim Davis
Staff Writer, CNET News.com
Published: November 18, 1996, 1:30 PM PST

I'll just skip to the good part...

"LAS VEGAS, Nevada--Intel (INTC) CEO and President Andrew Grove is still paranoid but instead of seeking therapy, he plans to use his paranoia to fuel the next generation of computer and processor technology, according to his keynote speech today at fall Comdex '96."

and

"By 2011, he projected, there will be a 200-fold increase in processing power on Intel processors, chips that will by then integrate 1 billion transistors and run at dizzying speeds of 10 GHz and 100,000 mips (millions of instructions per second). This compares with today's 4-million transistor processors running at 200 MHz and approximately 200 mips."

http://news.com.com/2100-1001-247914.html?legacy=c...

So, someone out there has married the number "10" to "2005" instead of "2011." Considering the willingness of most people to do any sort of fact checking whatsoever. I know that Cnet doesn't have a great reputation, but I can guess that not many sites have stories running back to 1996. For having made a hipshot more than a decade out, the claim of 10Ghz by 2011 sounds reasonable, especially considering the potential for the dubious marketing of multi-core chips.

So. 10 Ghz by 2005 is forum gossip that has been proven false. 10 Ghz by 2011 is official gossip that will probably turn out to be true...

Parent

RE: Supporting Links/Clarification

By masher2

(blog) on 7/11/2006 11:34:12 AM , Rating: 2

> "The nugget we've all been waiting for..."

Bravo, sir. Bravo.

Parent

RE: 32 Cores by 2010

By dilz on 7/11/2006 2:11:11 AM , Rating: 1

Apparently, even one of the editors at the venerable "Anandtech" was aware of this. Were you aware that Anandtech also has a search function? Put in "10 ghz" and you're served with this awesome quote:

quote:
"This reasoning explains why all of sudden Intel marketing does not talk anymore about 10 GHz CPUs, but about the "era of thread parallelism"."

http://www.anandtech.com/showdoc.aspx?i=2343&p=4

So, no matter who you want to blame/credit with having said it, apparently it was said and people accurately reported on it - at the time. (Whether it was Ace's, Anandtech, Tom's, or my favorite, or that elusive "Joe Sixpack" who rarely vists us in these forums.)

It's called conjecture. Every site stands accused of peddling gossip. That's what high-speed communications have allowed us to do.

In other news, a new day is starting in another part of the world. I could go on, but I'll stop. *YAWN*

Parent

RE: 32 Cores by 2010

By Knish on 7/11/2006 2:36:09 AM , Rating: 2

Not to threadspam, but Johan (author of the AT article) is also the founder of Aces hardware. He's probably the guy that started the original conjecture.

Parent

RE: 32 Cores by 2010

By deeznuts on 7/11/2006 3:15:13 AM , Rating: 3

Um, you didn't actually prove/disprove anything buddy. Nobody is doubting Intel has mentioned 10GHz parts, even I said so in my post. I would like to see where Intel had claimed, and people reported, that 10GHz procs would be here by 2005. Your article says Intel had stopped talking about it. No mention of 2005, and that article was posted in 2005.

And yes, I am quite "aware" Anandtech has a search function (not sure if your post was condescending but I'll give you the benefit of the doubt). I just choose to add links that add to the argument. This was interesting:
http://www.anandtech.com/showdoc.aspx?i=1379&p=6

quote:
Realistically speaking, we should be able to see NetBurst based processors reach somewhere between 8 – 10GHz in the next five years before the architecture is replaced yet again. Reaching 2GHz isn’t much of a milestone, however reaching 8 – 10GHz begins to make things much more exciting than they are today. Obviously this 8 – 10GHz clock range would be based on Intel’s 0.07-micron process that is forecasted to debut in 2005. These processors will run at less than 1 volt, 0.85v being the current estimate.

Looks like conjecture from Anand, maybe this started it all? It said the technology to get to 10GHz will be introduced in 2005 (65nm). Of course the predictions were all wrong, but that's besides the point.

Ok, how about a bone from the wonderful Anand search?
http://www.anandtech.com/showdoc.aspx?i=1584&p=5
Part of a proc.

Parent

RE: 32 Cores by 2010

By dilz on 7/11/2006 7:06:49 PM , Rating: 2

quote:
Um, you didn't actually prove/disprove anything buddy.

Actually, I was attempting to show that the Internet is infested by rumormongers. I admit defeat in accomplishing this. Anyway. I think I missed the sarcasm in one of the first few posts, and was on a mission from then on. To be honest, I thought I was writing more in favor of your position. Ahh.. the drawbacks of emotionless text!

<ENVY>
Deeznuts, since I have a last name that starts with "De," I tend to try and use that name. You're a lucky man to have secured such an alias!
</ENVY>

Parent

RE: 32 Cores by 2010

By freon on 7/11/2006 4:41:43 PM , Rating: 2

Intel never claimed 10Ghz BY 2005. Their claim was the possibility for 10Ghz in the next 5-10 years. That was in 2000.

The "10Ghz by 2005" rumor was started a few years ago when the Inq ran an article claiming to have a legit Intel roadmap showing 10Ghz released in 2005. This obviously didn't happen.

More recent leaked roadmaps show Conroe being replaced around 2008, then again with another new architecture around 2010. That seems to fit Intel's original estimate a little better.

Parent

RE: 32 Cores by 2010

By Questar on 7/11/2006 11:46:29 PM , Rating: 2

quote:
As someone said on Aces.

Intel also told us we'd have 10GHz processors by 2005"

Just thought I'd add that.

You forget that K9 was supposed to be out in 2005 running at 10Ghz.

Just thought I'd add that.

Parent

RE: 32 Cores by 2010

By Tyler 86 on 7/10/2006 9:58:11 PM , Rating: 2

They have self-contained water-cooling on the way.

Parent

RE: 32 Cores by 2010

By defter on 7/11/2006 4:31:53 AM , Rating: 1

Why the cooling would be a problem? Quad core chips will work fine on 65nm process.

First, this chip will be manufactured at 32nm process, that's two generation ahead of 65nm. In theory, it should reduce power by 3/4s compared to Intel's current process. Second, the chip will run at 2GHz according to the article, lower clockspeed means lower power consumption. Finally, the cores are likely to be more simple than current Conroe/Woodcrest cores. Keep in mind that this is targeted against Niagara which has a lot of small, simple integer cores.

One thing that is missing from the summary is that each core will support 4-way multithreading. Thus this CPU will be able to run 128 threads per socket.

Parent

its all about the process tech

By kattanna on 7/11/2006 12:19:56 PM , Rating: 2

there is no friggin way they are gonna have 32 cores on a single chip by then unless it is one MONSTER of a chip costing tens of thousands of dollars each.

they are gonna need at LEAST another 2, most likely 3, process shrinks to allow such a chip to be made and in economical numbers.

given current tech - quad cores
next shrink - 8 cores
next shrink - 16 cores
next - 32 cores

and considering each shrink usually takes 2 years to complete...thats 6 years

and 32 cores per physical chip..LOL um no...

besides process shrinkage they need they are gonna have to do some SERIOUS work on a completely new bus type..actually probably multiple buses in/out of the chip or the damn thing will starve to death.

and then there is simply need...

we have long needed multy core chips, and dual cores are a nice thing. quad cores? yeah they will be usefull, mostly for business. 8 cores...well now i think we are reaching the end of everyday practical use. will there be a true need and use..oh yeah..very high end

16..and the 32 cores...please. only the trully ultra high end, aka tens of thousands of chips made per year, are going to be needed, unlike the main stream hundreds of millions of chips

RE: its all about the process tech

By saratoga on 7/11/2006 2:53:01 PM , Rating: 2

You're overlooking a few things:

First, they only mention 768kb of cache per core, not the 2MB per core on Core 2.

Second, they don't say what type of Core is used. A P4 had an enormous core. Sun's T1 has tiny cores. So small they can fit 8 of them on one die using today's processes. In a massively parallel system like this, the dies probably are not symetric. Rahter there are probably 1-2 highend cores (OOOE, pipelined FPU, etc) and a much greater number of simplier cores that are more specialized and possibly even in order.

Parent

RE: its all about the process tech

By Motley on 7/13/2006 5:16:31 PM , Rating: 2

And nobody will ever need more than 640K of RAM either!

Short sighted. Very short sighted. Considering that todays high end video cards contain more pipelines than that. And they only have to render what you see. A CPU for games has to track every object in the game world. And most game worlds I know contain more than 32 objects. Ok, Pong contained 3. I guess it doesn't need 32 cores.

Parent