Intel aims for 32 cores by 2010

[Pinnacle]

"Keifer" - Intel's project name for many-core processors - will be surfacing in the 2009/2010 time frame and integrate 32 cores (128 threads total): Code-named "Gulftown," the chip will be manufactured in 32 nm and use eight processing nodes with four cores each. Every node will have direct access to one 3 MB on-die last level cache (LLC) unit and 512 KB L2 cache. There will be a total of 8 x 3 MB LLC slices that are connected by a ring architecture and represent a total of 24 MB LL cache.
While it is worth noting that Intel expects Keifer to result in a 15x performance jump over today's Xeon 5100 processors at just 2 GHz clock speed, it is especially interesting that Intel has begun investigating competitive products to estimate the performance requirements of its products in the 2010 time frame


Holly Molly!

http://www.tgdaily.com/2006/07/10/intel_32_core_processor/

[n00b 0f l337]

Moore's law....

[alexio]

Finally they are advancing as fast as GPU's have in the last few years :)

However don't forget that this is an Intel estimate. They also estimated 5GHz Precotts and 10Ghz succesors of that.

About half the real performance is not conservative for anything claimed by Intel that is more than 2 years away.

[n00b 0f l337]

They also estimated 5GHz Precotts and 10Ghz succesors of that.
Maybe they meant dual cores, 5 x 2 (lol this is a joke)

[alexio]

Maybe they meant dual cores, 5 x 2 (lol this is a joke)
5 Northwood A's in one package with a die schrink to 65nm. Intel's view on advancing technology some years ago :D

[Vapor]

And finally we see where there's room to be gained...:)

Out of curiosity though, how do they anticipate the memory market to scale? DDR2 is more than enough for 4 cores (assuming current mem-reliance), then DDR3 will be too much for 8 and but possibly too little for 16, but 32 cores? Anyway, things will be very fun it seems :)

[zabomb4163]

keep in mind tejas probably would have reached 10ghz were it not for the leakage problem.

[Vapor]

keep in mind tejas probably would have reached 10ghz were it not for the leakage problem.Which, incidentally, has been fixed. Not to mention more expensive processes that are WAY more than capable of achieving this.

Still though, 32 cores will be a good number by 2010....so close when you think about it :eek: starting to doubt the plausibility tbh now....

[shadowing]

I would honestly guess 2015. 2010 seems very hard to reach.

This is considering how we have only just reached quad core. I expect by 2009 that there would be up to 8 cores.

[thunderstruck!]

Which, incidentally, has been fixed. Not to mention more expensive processes that are WAY more than capable of achieving this.

Still though, 32 cores will be a good number by 2010....so close when you think about it :eek: starting to doubt the plausibility tbh now....
Exactly. Look where we were four years ago. Athlon XPs were coming out. In a four year span we've gone from single-core to dual-core, or possible quad-core if you count Kentsfield. Now, another four years will go by, and 32 cores will be the norm? Who neads 128 threads anyway? (Except for running Rosetta;) )

[zabomb4163]

Exactly. Look where we were four years ago. Athlon XPs were coming out. In a four year span we've gone from single-core to dual-core, or possible quad-core if you count Kentsfield. Now, another four years will go by, and 32 cores will be the norm? Who neads 128 threads anyway? (Except for running Rosetta;) )

NASA, DOD, any weather station, CGI workers, professional video editers, gamers, department of energy, ect ect. List goes on and on. As the power becomes availble more companies will make use of it. Remember the bill gates quote about who would ever want a personal computer? Or 500KB would be enough space or anyone.

[Shadowmage]

Not sure about gamers... it's hard to code 3 concurrent threads doing relatively equal work for games, much less 32 or 128 threads.

[Reznik Akime]

They really need to implement some kinda serial bus like HT for that kinda power. Think of the bandwith, man!

[thunderstruck!]

As the power becomes availble more companies will make use of it. Remember the bill gates quote about who would ever want a personal computer? Or 500KB would be enough space or anyone.
"Nobody will ever need more than 640k RAM!" -- Bill Gates, 1981

[gOJDO]

@Reznik Akime
They will implement CSI in the next architecture, no need to worry about.

@shadowing
In 2015 you can expect much more than that.
If they can make quadcore on 65nm with 8MB of L2(which is takng more than 60% of the die space), than with the same size of silicon they can make 8 quadcores with 512kB L2 & 3MB of LLC(L3) using 32nm(4 times more effective die space).
For 32nm they are going to use tri-gate transistors which are the most fastest and lowest power consuming. Taking in mind diferent P & N states, the power consumation and the produced heat will be reduced further so we can expect high freqfency operation of the cores.

[zakelwe]

Fine for the server market but for desktop I am not sure whether it will be needed ( apart from DC of course ). 4 and 8 fine but 32 ?

Regards

Andy

[gOJDO]

In 2010 almost all the software will be multithreaded. No wonder, there will be use of more cores. Also, don't forget there are techniques to boost singlethread performance of non linear threads. http://www.intel.com/technology/magazine/research/speculative-threading-1205.htm

[alexio]

The pipelines in a GPU are basicly seperate cores. I guess they said the same about how hard it was to program in the days the first multi-pipeline core was introduced. I guess it will work out with CPU's as it did with GPU's. Programmers will have to if they want to compete with other programmers, certainly in the case of gamedevs.

[3NZ0]

That thing is going to EAT bandwidth like no other and without huge amounts of cache its going to need huge amounts of memory bandwidth, bi-quad data rate in ddr 5/6. :p:


but we dont have conroe yet....lets get our prioritys right :(

[PkG.1337]

Go beyond Moore's law, damn it intel.

[asmodean]

"Nobody will ever need more than 640k RAM!" -- Bill Gates, 1981

wrong.

http://en.wikiquote.org/wiki/Bill_Gates#Wrongly_Attributed

[Thorry]

Well I think the 32nm statement by 2010 is pretty bold to say the least.

ASML (and other chip-producing-machine-producers) have often stated it is impossible with the current techniques to go beyond 34~35nm and they will prolly be stuck at 37nm for the next couple of years.

A well 2010 is 3.5 years away... we will see what the brilliant guys come up with..

[gOJDO]

That thing is going to EAT bandwidth like no other and without huge amounts of cache its going to need huge amounts of memory bandwidth, bi-quad data rate in ddr 5/6. :p:


but we dont have conroe yet....lets get our prioritys right :(
Yes it will eat bandwidth, but the chip will have 24MB of LLC(last level cache) shared. I think that quad channel DDR3 will fead the beast succesfully.


Well I think the 32nm statement by 2010 is pretty bold to say the least.

ASML (and other chip-producing-machine-producers) have often stated it is impossible with the current techniques to go beyond 34~35nm and they will prolly be stuck at 37nm for the next couple of years.

A well 2010 is 3.5 years away... we will see what the brilliant guys come up with..
With tri-gate transistors, there will be no problems achieving 22nm. Intell are the first who will start using tri-gate transistors starting with 45nm, the next year.
http://www.intel.com/technology/silicon/tri-gate-demonstrated.htm
http://www.intel.com/pressroom/archive/releases/20020919tech.htm
http://www.intel.com/technology/magazine/silicon/si07031.pdf

[nn_step]

Dear lord here we go again.

[XS Janus]

:surf: :surf: And all that + 2 architectures better than Conroe (according to their "new architecture every 2 years" plan), each of which will probably be 10-15% better clock per clock than Conroe. :surf: :surf:
________
Free amateur stream (http://www.:banana::banana::banana::banana:tube.com/)

[.sentinel]

Actually they are going to run at around 1ghz each.

[Thorry]

With tri-gate transistors, there will be no problems achieving 22nm. Intell are the first who will start using tri-gate transistors starting with 45nm, the next year.
http://www.intel.com/technology/silicon/tri-gate-demonstrated.htm
http://www.intel.com/pressroom/archive/releases/20020919tech.htm
http://www.intel.com/technology/magazine/silicon/si07031.pdf

Well maybe it will help a bit, but that won't get you into the 22nm range. It's still not possible to get beyond the 32nm with this technique. The features however will be equal to what the old techniques would have yielded at 22nm, but the production itself is simply not possible with the current lithographic techniques.

It doesn't matter what you want to put on the wafer, size is all that matters.

[flutie98]

Well maybe it will help a bit, but that won't get you into the 22nm range. It's still not possible to get beyond the 32nm with this technique. The features however will be equal to what the old techniques would have yielded at 22nm, but the production itself is simply not possible with the current lithographic techniques.

It doesn't matter what you want to put on the wafer, size is all that matters.


if only the term self scattering didnt exist.... the possibilities

[aMp]

There's some Gates-libel going on here. The earlier quote about nobody needing a personal computer
was by Ken Olsen (http://www.snopes.com/quotes/kenolsen.asp), co-founder of DEC, not Bill. Actual quote, from 1977: "There is no reason for any individual to have a computer in his home."

That's twice I've stood up for Bill/MS in the last year. I must be getting soft in my old age.

[nn_step]

Well maybe it will help a bit, but that won't get you into the 22nm range. It's still not possible to get beyond the 32nm with this technique. The features however will be equal to what the old techniques would have yielded at 22nm, but the production itself is simply not possible with the current lithographic techniques.

It doesn't matter what you want to put on the wafer, size is all that matters.
if you are talking about production efficiency. that is very true. However.
A brilliant 130nm Design is competitive with poor 65nm designs. Although the process Helps greatly. It isn't everything.

[gOJDO]

if you are talking about production efficiency. that is very true. However.
A brilliant 130nm Design is competitive with poor 65nm designs. Although the process Helps greatly. It isn't everything.
better production processes are primary invnented becouse of the shrink of the die. It is becouse of the lower production cost of the chip and higher yields per waffer.

[gOJDO]

Well maybe it will help a bit, but that won't get you into the 22nm range. It's still not possible to get beyond the 32nm with this technique. The features however will be equal to what the old techniques would have yielded at 22nm, but the production itself is simply not possible with the current lithographic techniques.

It doesn't matter what you want to put on the wafer, size is all that matters.
With tri-gate transistors occupy die space in 3D. 45nm tri-gate process development is finished and waiting to be used. Intel are allready working on 32nm and 22nm.

[Thorry]

better production processes are primary invnented becouse of the shrink of the die. It is becouse of the lower production cost of the chip and higher yields per waffer.

This is a common misconception:

The die shrink is only being done to reduce power (and thus reduce leakage and thus increase switching speeds and thus increasing overall speed allowing for higher clockspeeds).

The yields actually have nothing to do with this.
When going to a smaller process the yield don't change, if anything the yields get lower. (Because it is a new technique).

The amount of cores per wafer increases, but the amount of errors per wafer also increases (that's simple statistics).

It also isn't less expensive to produce, it takes a lot of effort to switch to a new process. Altough a single core is cheaper to produce, when throwing in the R&D and effort that went into switching over to the new process (including some new hardware that Intel has to buy over at ASML) the actual price per core is pretty much the same (although it will be a faster core so the actual added value increases making a lower retail price possible).

So technically switching to a smaller procede doesn't make it cheaper even though the retail value can become cheaper (but this is up to the manufactor, if AMD didn't exist I doubt we will see cheaper CPUs when going to a smaller procede. It may even be more expensive in the short term).


With tri-gate transistors occupy die space in 3D. 45nm tri-gate process development is finished and waiting to be used. Intel are allready working on 32nm and 22nm.

I understand tri-gate transistors, actually read a very thorough article about it in 2003.

However: Intel can design all they want, hell you can prolly design a 12nm chip and make it do all you ever wanted.

But it's guys like ASML and the sorts who have to build the machines that process the wafers into chips. If they can't get the smaller process together Intel isn't doing anything with the designs.

I have no doubt that ASML (and all the other brainy guys over the world) get it right and find some way to go beyond 32nm. But stating it will be here in 2010 is doubtfull.

Also you can't simply keep on going smaller and smaller, at 22nm you can actually get into trouble due to quantum field effects. They will have to think of some way (and 3D seems the way to go) to get more out of the same.

However, 32nm is still 32nm nomatter how many directions you travel in ;)

[Qkjhfhaiguihfma]

even at 32nm, won't 32 cores on 1 die be larger than what we are used to seeing?

[lawrywild]

I'd much rather have 8 faster cores than 32 slower cores..

[XS Janus]

even at 32nm, won't 32 cores on 1 die be larger than what we are used to seeing?

If this holds up, I bet will see simpler cores with efficient designs and therefore dies of about the same size like today's.
Doesn't really matter how big it is if it doesn't produce heat anyway... (remember all that Intel talk about ultra, ultra low wattage cores ;) ) Who knows... Intel Brainstorming like this is Exciting nonetheless :woot:
________
UTG SPECIAL OPERATIONS UNIVERSAL TACTICAL (http://airsoft-shop.info/p/utg-special-operations-universal-tactical/)

[nn_step]

even at 32nm, won't 32 cores on 1 die be larger than what we are used to seeing?
It all depends on the individual core complexity.
For example lets just say the cores are as complex as the 4004 processor with 2,300 transistors apiece. then Even 2,000 cores would make the die about the size of a Pentium 4 Processor, single core... Not even dual core. (Which has 55 million transistors if I remember correctly, dual has twice that) and the Pentium 4E has 125 million Transistors (mostly Cache). So once again. It is all about how complex the cores are.:D

[realsmasher]

somebody got skills in PS to create a logo of that monster ?

you know, such fake stuff like the "core 3 octo" picture.

just like this but more nice and maybe other color.

http://img147.imageshack.us/img147/9550/core67mn.jpg

I'd really love it.

also, who knows how to name ?

duo, quad, octo but what comes after ? hexa ? then ?

this is the most important question now :D

[Qkjhfhaiguihfma]

core triacontadi

[[cTx]Philosophy]

wrong.

http://en.wikiquote.org/wiki/Bill_Gates#Wrongly_Attributed

Man you beat me to it :(
Good catch though...
Gates FTW!!!!

Oh about the procesors, umm yah there gonna be fast :) :slapass:

[[XC] leviathan18]

they said 10ghz too

[Thorry]

somebody got skills in PS to create a logo of that monster ?

you know, such fake stuff like the "core 3 octo" picture.

just like this but more nice and maybe other color.

http://img147.imageshack.us/img147/9550/core67mn.jpg

I'd really love it.

also, who knows how to name ?

duo, quad, octo but what comes after ? hexa ? then ?

this is the most important question now :D

Greek:

Ena = 1
Dio = 2
Tessera = 4
Octo = 8
Hexa = 6 (or actually hexi)
Hexadeca = 16 (or actually hexideca)

Latin:

Duo = 2
Quad = 4
Octo = 8
Sedecim = 16
Vigintiquad = 24 (actually vigintiquattuor)
Trigintaduo = 32
Sexagintaquad = 64 (actually sexagintaquattuor)
Centumduodetriginta = 128

Intel's gonna have a hard time naming it ;)

[nn_step]

Somehow I think sexagintaquattuor might catch on :p:

[realsmasher]

pretty nice names :banana:

i think that's the reason why they won't continuo with name = core count

it even hardly fits on the logo ! :eek: