Intel 32nm Westmere Desktop Processor Roadmap Exposed

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Nemesis 1

Lifer
Dec 30, 2006
11,366
2
0
Originally posted by: Idontcare
Originally posted by: Cogman
BTW, will the graphics chip be manufactured at the same node size as the rest of the processor?

For Arrandale (and Clarkdale) the GPU is MCM'ed with 45nm GPU parts.

When the Sandy Bridge refresh comes out the following year (Q4 2010) there is a Cougar Point chip which replaces Arrandale and Clarkdale.

Cougar Point is supposed to be a truly integrated GPU/CPU (monolithic) all at 32nm with Sandy Bridge core logic.

What is not known, as far as I am aware, is whether this integrated GPU in Cougar Point will be Larrabee based or be another iteration of the existing GPU architecture that Intel is fielding present day.


My bet would be An 8core larrabee maybe more. When one looks at Sandy . One must first look at AVX. That will tell you were Intel is going.

http://www.uberpulse.com/us/20..._at_next_idf_video.php

 

magreen

Golden Member
Dec 27, 2006
1,309
1
81
I read the AT 32nm article today. Maybe I'm reading those weird Intel slides wrong (they're hard to read), but I don't see a Nehalem-based server any time soon, as in not before end of year 2009. Where the heck are they?

I thought the whole point of Nehalem was to take back the server space, and the minor improvements on desktop were gravy. So where are they? Last time I asked, idc you said validation. But a whole year of validation? The desktop part has been out for months.
 

aigomorla

CPU, Cases&Cooling Mod PC Gaming Mod Elite Member
Super Moderator
Sep 28, 2005
20,839
3,174
126
Originally posted by: magreen
I read the AT 32nm article today. Maybe I'm reading those weird Intel slides wrong (they're hard to read), but I don't see a Nehalem-based server any time soon, as in not before end of year 2009. Where the heck are they?

I thought the whole point of Nehalem was to take back the server space, and the minor improvements on desktop were gravy. So where are they? Last time I asked, idc you said validation. But a whole year of validation? The desktop part has been out for months.

i was trying to point that out.

the sides are old and wrong.

according to those slides, we should already have i5's, but according to that roadmap, 32nm i5's are around the corner when the 45nm's were just previewed not too long ago.

either that or intel picked up a bad habit from AMD called DELAYING BS.

Originally posted by: Ocguy31
Originally posted by: angry hampster
Seems like the Nahelem early adopters really got dicked on this one. .

When do early adopters not get dicked?

Bring on 32nm!

Well lets see.. when G0's first came out, they all did pretty much 3.6
The late migraters to G0 all got dicked @ 3.2

So its the otherway around in this case.

Earily batch E6600 also clocked better then later batch E6600. so its not always the case.
 
Oct 19, 2006
194
1
81
Originally posted by: PCTC2
Originally posted by: BTRY B 529th FA BN
Just been reading the responses, so forgive my asking; 1366 socket still?

Yes, for Gulftown, but the Performance/Mainstream segment (Lynnfield, Clarkdale) are Skt1156.

Originally posted by: angry hampster
Seems like the Nahelem early adopters really got dicked on this one. I may be in for one of those new dual core/quad thread Clarkdales.

Uhh. I don't think I"m dicked for one. If you happen to notice, the Extreme segment of the market will remain the X58 and Skt 1366. Sure, Lynnfield with an integrated PCIe controller and IMC looks nice, but I sure like my i7. Plus, Skt 1366 on the X58 chipset gets Gulftown w/ 6 cores and 12 threads on the X58. I'm pretty sure I'll just plop one of those into my X58 and I'm not going to feel dicked. At all. Have fun with your quad-core when I have my sex-core.

Early adopters almost always get punished by intel. Remember socket 423 and socket 478 wasn't to great either. But As the anandtech article said,you might need a new motherboard to run the six core Gultown. Remember the early adopters of socket 775? Hell most of the second gen 775 boards won't run with a 45nm C2Q or C2D.

What puzzles me is the lack of a 32nm Quad? How many people are really going to buy a 6 core beast? So your mid to enthusiast customers will be buying old 45nm stock. Seems pretty crazy to me. I almost don't believe it.
 

ilkhan

Golden Member
Jul 21, 2006
1,117
1
0
we'll probably see quad core 32nm for desktops, but it'll be a LGA1156 part, the sixcore will probably be LGA1366 only.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: anarchyreigns
Originally posted by: dmens
But As the anandtech article said,you might need a new motherboard to run the six core Gultown

No, absolutely not.

And you know that how?

;) You need to spend some more time around here friend, then you might know who you are talking to and why you should know they know exactly what they are talking about ;)

hint: dmens sees them everyday
 

dmens

Platinum Member
Mar 18, 2005
2,271
917
136
Originally posted by: anarchyreigns
And you know that how?

because the two designs are supposed to be socket compatible and should only require a BIOS update.

also, it would be really retarded if it were not the case.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: aigomorla
i was trying to point that out.

the sides are old and wrong.

according to those slides, we should already have i5's, but according to that roadmap, 32nm i5's are around the corner when the 45nm's were just previewed not too long ago.

either that or intel picked up a bad habit from AMD called DELAYING BS.

aigo these slides came directly from Intel and were released to the media (anandtech, hot hardware, tooms, etc) at a coordinated press release and conference held on Tuesday.

That are as authentic and as close to the latest public roadmaps as it gets.

If the information contained in the slides does not jive with your expectations (or anyone else's) then the purpose of the conference is fulfilled, your expectations are to be re-aligned with the new/planned reality as demonstrated in the slides.

Once caveat to reading release timelines to those vague marketing slides where the x-axis appears to be linear in time but there are no sub-markings at less than annual intervals...buyer gets what they payed for when assuming something is to be released 1H or 2Q vs 2H or 4Q from those slides.

This slide for example, the reader could easily be convinced that Intel means to imply the start of 2009 calendar year lines up with a vertical line that would hit right around where the 2009 label is. For all we know they centered the label 2009 within the actual 2009 time-range on the chart, so all the stuff we think should be here now isn't actually schedule for release until June or July.

No mystery, just misinterpretation of what the timeline means.

At any rate my point is this - don't try and scrutinize the slides looking for reasons to call shens based on what you think you know from past roadmaps or leaked conversations. Everything got changed up right after Christmas once early 32nm yields came in so well and the financial situation of Q4 collapsed so markedly.

The coach just thru out the playbook for the game and called in plan B. Take plan B at face-value, forget what you thought you knew was going to happen, and progress from here.

Originally posted by: magreen
I read the AT 32nm article today. Maybe I'm reading those weird Intel slides wrong (they're hard to read), but I don't see a Nehalem-based server any time soon, as in not before end of year 2009. Where the heck are they?

I thought the whole point of Nehalem was to take back the server space, and the minor improvements on desktop were gravy. So where are they? Last time I asked, idc you said validation. But a whole year of validation? The desktop part has been out for months.

I think you might be misreading the charts. Remember Nehelem-EP = Nehalem Xeon for gainestown (1S and 2S). Nehalem-EX = Beckton = 4S platforms (think dunnington)

Nehalem-EP comes out 1H/09

http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=4

http://images.anandtech.com/re.../enterpriseroadmap.jpg
 

nitromullet

Diamond Member
Jan 7, 2004
9,031
36
91
Originally posted by: ilkhan
lynnfield (45nm, 4c/8t, no GPU) will be LGA1156
clarksfield (45nm, 4c/8t, no GPU, mobile) will be 989


gulftown (32nm, 6c/12t, no GPU) will be LGA1366
unknown/non-revealed (32nm, 4c/8t, no GPU) will be LGA1156
clarksdale (32nm, 2c/4t, GPU) will be LGA1156
arrandale (32nm, 2c/4t, GPU, mobile) will be 989

up until mid 2010 (when we get near/to sandy bridge) looks like
high end desktop gulftown
quad desktop lynnfield
dual desktop clarksdale
quad mobile clarksfield
dual mobile arrandale

Id expect a quad 32nm product for the desktop, but we havent heard anything about that yet.

That is the wild card really... it seems like a decision Intel will make when the time comes.

Right now the road maps don't reflect that:

Lynnfield - 45nm, 4c/8t, LGA1156
http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=4

...nothing 32nm, 4c/8t, LGA 1156
http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=5

I'm guessing that Intel would prefer to keep LGA 1366 as the 4+ core socket going forward after the 32nm switch, but I imagine they would suffer some serious backlash from the 'community' - especially if LGA 1366 boards continue to cost around $200+. Plus, not having an affordable quad-core for the mainstream would probably give AMD a free pass to fill that market niche.

Originally posted by: ilkhan
we'll probably see quad core 32nm for desktops, but it'll be a LGA1156 part, the sixcore will probably be LGA1366 only.

Agreed.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: nitromullet
Originally posted by: ilkhan
lynnfield (45nm, 4c/8t, no GPU) will be LGA1156
clarksfield (45nm, 4c/8t, no GPU, mobile) will be 989


gulftown (32nm, 6c/12t, no GPU) will be LGA1366
unknown/non-revealed (32nm, 4c/8t, no GPU) will be LGA1156
clarksdale (32nm, 2c/4t, GPU) will be LGA1156
arrandale (32nm, 2c/4t, GPU, mobile) will be 989

up until mid 2010 (when we get near/to sandy bridge) looks like
high end desktop gulftown
quad desktop lynnfield
dual desktop clarksdale
quad mobile clarksfield
dual mobile arrandale

Id expect a quad 32nm product for the desktop, but we havent heard anything about that yet.

That is the wild card really... it seems like a decision Intel will make when the time comes.

Right now the road maps don't reflect that:

Lynnfield - 45nm, 4c/8t, LGA1156
http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=4

...nothing 32nm, 4c/8t, LGA 1156
http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=5

I'm guessing that Intel would prefer to keep LGA 1366 as the 4+ core socket going forward after the 32nm switch, but I imagine they would suffer some serious backlash from the 'community' - especially if LGA 1366 boards continue to cost around $200+. Plus, not having an affordable quad-core for the mainstream would probably give AMD a free pass to fill that market niche.

Originally posted by: ilkhan
we'll probably see quad core 32nm for desktops, but it'll be a LGA1156 part, the sixcore will probably be LGA1366 only.

Agreed.

Everything in the 4c/8t segment changed once AMD released 45nm Phenom II and Shanghai Operton.

Once Intel knew what their westmere chips were going to be up against for the next 1-2 yrs it became obvious that at the price segments and clockspeeds AMD would be bringing to the market a 2c/4t westmere chip at the clockspeeds they expect to deliver will be more than enough to compete price/performance with AMD while delivering the gross margins and volume sales Intel wants to see from their mainstream desktop offerings.

The only way we'll see 4c/8t on mainstream market is if AMD's Sao Paolo 6core chip gets released to the desktop segment and competes with the performance of Clarkdale at whatever clockspeeds Clarkdale comfortably tops out at.

And why go 6c/12t for enthusiast segment? Because you can't overclock the number of cores. You can overclock the clockspeed, buy a low-clocked mainstream 4c/8t chip and overclock it to get the same performance of a high-end enthusiast 4/8t bloomfield...but if the mainstream chips are only 2c/4t then it doesn't matter how high you overclock it you still aren't going to come close to the performance people are getting with their 6c/12t rigs.

In other words, more than one way to skin a cat. Let mainstream people continue to invent cleverer and cleverer ways to overclock mainstream chips, Intel won't bother to stop you from trying. But your not getting a 4c or a 6c mainstream chip to OC anyways, so your performance will not eclipse the enthusiast market anymore.
 

magreen

Golden Member
Dec 27, 2006
1,309
1
81
Originally posted by: Idontcare
Originally posted by: magreen
I read the AT 32nm article today. Maybe I'm reading those weird Intel slides wrong (they're hard to read), but I don't see a Nehalem-based server any time soon, as in not before end of year 2009. Where the heck are they?

I thought the whole point of Nehalem was to take back the server space, and the minor improvements on desktop were gravy. So where are they? Last time I asked, idc you said validation. But a whole year of validation? The desktop part has been out for months.

I think you might be misreading the charts. Remember Nehelem-EP = Nehalem Xeon for gainestown (1S and 2S). Nehalem-EX = Beckton = 4S platforms (think dunnington)

Nehalem-EP comes out 1H/09

http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=4

http://images.anandtech.com/re.../enterpriseroadmap.jpg

Ok, right, I remember now that the 2P server chips are coming out the middle of this year. But isn't the place that Intel needed the most rescuing the 4P server space? That's where AMD's HTT was eating the Xeons and their FSB alive, no?
 

nitromullet

Diamond Member
Jan 7, 2004
9,031
36
91
Originally posted by: Idontcare
Everything in the 4c/8t segment changed once AMD released 45nm Phenom II and Shanghai Operton.

Once Intel knew what their westmere chips were going to be up against for the next 1-2 yrs it became obvious that at the price segments and clockspeeds AMD would be bringing to the market a 2c/4t westmere chip at the clockspeeds they expect to deliver will be more than enough to compete price/performance with AMD while delivering the gross margins and volume sales Intel wants to see from their mainstream desktop offerings.

The only way we'll see 4c/8t on mainstream market is if AMD's Sao Paolo 6core chip gets released to the desktop segment and competes with the performance of Clarkdale at whatever clockspeeds Clarkdale comfortably tops out at.

And why go 6c/12t for enthusiast segment? Because you can't overclock the number of cores. You can overclock the clockspeed, buy a low-clocked mainstream 4c/8t chip and overclock it to get the same performance of a high-end enthusiast 4/8t bloomfield...but if the mainstream chips are only 2c/4t then it doesn't matter how high you overclock it you still aren't going to come close to the performance people are getting with their 6c/12t rigs.

In other words, more than one way to skin a cat. Let mainstream people continue to invent cleverer and cleverer ways to overclock mainstream chips, Intel won't bother to stop you from trying. But your not getting a 4c or a 6c mainstream chip to OC anyways, so your performance will not eclipse the enthusiast market anymore.

That actually makes a lot of sense. Hopefully, there will still be good quad core mid-high options. I'm thinking as long as there is still a $200-300 quad core option, I'll be happy. Aside from that, X58 has the huge advantage of being both CF and SLI compatible in most cases. I'm assuming that P55 either won't be SLI qualified by NVIDIA, or will only have enough PCIe lanes to do 8x/8x PCIe. It's actually good to see that X58/LGA1366 looks like it will have a long life expectancy.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: magreen
Originally posted by: Idontcare
Originally posted by: magreen
I read the AT 32nm article today. Maybe I'm reading those weird Intel slides wrong (they're hard to read), but I don't see a Nehalem-based server any time soon, as in not before end of year 2009. Where the heck are they?

I thought the whole point of Nehalem was to take back the server space, and the minor improvements on desktop were gravy. So where are they? Last time I asked, idc you said validation. But a whole year of validation? The desktop part has been out for months.

I think you might be misreading the charts. Remember Nehelem-EP = Nehalem Xeon for gainestown (1S and 2S). Nehalem-EX = Beckton = 4S platforms (think dunnington)

Nehalem-EP comes out 1H/09

http://www.anandtech.com/cpuch...howdoc.aspx?i=3513&p=4

http://images.anandtech.com/re.../enterpriseroadmap.jpg

Ok, right, I remember now that the 2P server chips are coming out the middle of this year. But isn't the place that Intel needed the most rescuing the 4P server space? That's where AMD's HTT was eating the Xeons and their FSB alive, no?

There's "oh noes our chips are being eaten alive" and then there's "darn, we aren't making as much penetration into the $100M/year 4S market space as we'd like, guess we'll just have to survive on our paltry remaining $9B business in the other segments".

AMD is in the death throes of "oh noes! our desktop segment, our largest dollar-volume segment, is being slaughtered by Intel". Sure they might have 90% of the 4S market, 90% of $100M is how much in gross sales per year?

Intel wants that business because they are entitled to it in the sense they have the architecture now that enables them to own that market space if they set the pricepoints accordingly. But they are hardly hurting by taking their sweet time getting it there.

Consider that Beckton is 700mm^2 die. Even with mature production yields they are probably lucky to get 10-15 sellable die off a wafer at this time. Sure they are going to be priced so high that even at 15 chips per wafer they still make money, but with those numbers against you it doesn't hurt to spend an extra quarter or two improving yields another 5 die per wafer while ensuring validation is bar-none stellar at time of release.
 

pm

Elite Member Mobile Devices
Jan 25, 2000
7,419
22
81
Consider that Beckton is 700mm^2 die. Even with mature production yields they are probably lucky to get 10-15 sellable die off a wafer at this time.
That could be right - if you don't factor in cache redundancy.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: pm
Consider that Beckton is 700mm^2 die. Even with mature production yields they are probably lucky to get 10-15 sellable die off a wafer at this time.
That could be right - if you don't factor in cache redundancy.

Yeah I remember with our SUN chips that were *tiny* compared to 700mm^2 behemoths we started "volume" production when we hit 5 NUBs (net units built) per wafer.

As we improved yields and the NUBs increased we just backed off on total wafer starts per day for that particular device. The accounting just seemed like it should have been horrific, but SUN was happy to buy wafers with 5 NUBs. Of course one has to take into account SUN's fiscal wherewithal in the business acumen department though...

But when you are popping chips out that someone will pay $2.5k for, you actually don't need that many of them to hit break-even on your production costs. I would love to see a Beckton wafer with my own eye after having my expectations of what "big die" means having been calibrated with those measly 300-400mm^2 SUN chips. I think an OMFG might slip out on such an occasion.
 

ilkhan

Golden Member
Jul 21, 2006
1,117
1
0
Originally posted by: nitromulletAside from that, X58 has the huge advantage of being both CF and SLI compatible in most cases. I'm assuming that P55 either won't be SLI qualified by NVIDIA, or will only have enough PCIe lanes to do 8x/8x PCIe. It's actually good to see that X58/LGA1366 looks like it will have a long life expectancy.
X58 is connected to 32 (actually, 40) PCI-E lanes.
LGA1156 has 16 lanes built into the CPU. x16(single) or x8/x8 max, depending on which version of the x5x chipset is being used. Maybe its 20 lanes, so give some extras for other devices....but I think its only 16. Maybe from ibix peak via DMI...

Wow, Idontcare. 300-400 possible chips per wafer (300mm wafer and about 350mm^2 each, if my math is right) and they were happy with *5*?!
 

pm

Elite Member Mobile Devices
Jan 25, 2000
7,419
22
81
I think an OMFG might slip out on such an occasion.

The CPU that I work on is bigger than Beckton. :) And yeah, a common reaction on seeing a bare die from the CPU that I work on is "wow, that thing is huge." Which is always ironic when describing something about the size of a large toenail. :)
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Are you sure you aren't confusing with Tukwila on the 700mm2 die size??

ilkhan: X58 has 36 PCI-Express 2.0 lanes not 32.

EDIT: The die size looks somewhere around 600mm2.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: IntelUser2000
Are you sure you aren't confusing with Tukwila on the 700mm2 die size??

ilkhan: X58 has 36 PCI-Express 2.0 lanes not 32.

I'm intentionally being generous with my die-size estimates in homage to the fact that really anything that large need not be quantified to three significant digits. :D

Yes Beckton is more like 580-600mm^2 as it is 2x Nehalem (2x263'ish = 526'ish) plus another 8MB L3$ gets shoved onto the chip...so 600mm^2.

Needless to say the challenges with this chip are more in the league of comparably larger 700mm^2 tukwila, 503mm^2 Dunnington, and 576mm^2 GT200 chips than in the league of anything <300mm^2 like Nehalem, Phenom, Niagara, Power5, etc.

Originally posted by: ilkhan
Wow, Idontcare. 300-400 possible chips per wafer (300mm wafer and about 350mm^2 each, if my math is right) and they were happy with *5*?!

Well...I should state it more precisely as "we were happy" as in TI the foundry. But yeah that is why you don't talk yields in percentage with large die like that, saying you have 1.8% yield is depressing and doesn't communicate what you really need to know for wafer start planning. Saying you get 5 NUBs (five shipped chips) per wafer sounds a little less depressing, but does communicate what planning needs to know because they are trying to fill chip-out orders, not wafer orders.

At any rate those are time-zero production yields at qual, where hitting reliability is the rate-limiting step to shipping product, not volume ramp issues. The fab then spends the next year thereafter continuously improving yields. Once maturity is reached then you'll see so many nubs coming off a wafer (>100, or 40-50% yield) that it will have long since been converted to %-yield speak again. This transition (for us anyway) usually occurred around the 8-10% yield levels, above 10% yield we spoke of yield in percentages, below 10% we spoke in terms of nubs.
 

Idontcare

Elite Member
Oct 10, 1999
21,118
58
91
Originally posted by: IntelUser2000
Probably not the same thing as seeing in person, but the wafer at the picture below is apparently containing Nehalem-EX.

http://www.heise.de/resale/IDF...Plaene--/zoom/114504/0

The L3 is actually extremely small in comparison to the core in Nehalems. Per MB die size is only 6mm2.

Thanks for the link, I had seen that photo before but for the life of me could not track it down afterwards :beer:

Yeah it's huge, look at it relation to Pat's fingernails.

Makes me wonder just how freaken big pm's toes must be if he thinks those monsters are "about the size of a large toenail"...:laugh: Yeah, maybe Shaquille O'Neal's toenails :p
 

Denithor

Diamond Member
Apr 11, 2004
6,300
23
81
So no 4/8 32nm chips for LGA1366?

Because the LGA1156 isn't as interesting due to the reduced PCIe lane count. Do they have SLI/CF enabled for LGA1156? Even at the lower bandwidth most cards should run ok, I suppose.

And that 2/4 32nm, while interesting, just won't have enough horsepower going forward for games. Today's games can max two cores already so there's no advantage from hyperthreading on a dual-core (for gaming).
 

nyker96

Diamond Member
Apr 19, 2005
5,630
2
81
if 1156 is as intel planed to restrict overclocking i'd probably take a pass on it. good article though. 1366 will probably still be the next buy unless amd has a nice pII for am3 platform to compete in the ddr3 domain.