What will AMD do between now and bulldozer?

[KingstonU]

There is a long time between now and Bulldozer's expected tape out and production in late 2010 and product availability by 2011. That's 14+ months!!

The current spin is (RB-C2? CPU-Z) has helped Phenom a lot but has been in production for how long now? It seems to have a thermal wall since the 965 is rated at 140W TDP. Is there at least new-spin of Deneb scheduled to take it 3.6-4.0+ GHz at stock and lift the overclocking ceiling of 4.0GHz to something like 4.5+GHz?

The only other things on the horizon are desktop version of the 6-core Istanbul and Opteron's 12-core Magny-Cours. Then I guess HD6000 series GPU's by 2H 2010.

 

[WaitingForNehalem]

They'll stay in the low end market.

 

[RussianSensation]

Just like during Athlon XP days, AMD will compete in the lower price brackets. Currently AMD has a better processor than Intel for every price segment < $130 imo.

Right now I can only see 2 paths for a new system: AM3 or i5 750+. Socket 775 is a horrible 'investment' from DDR2 standpoint and price/performance. AMD is offering an extra core at the same price. Q9550 is ridiculously overpriced. AMD's problem has always been marketing their performance to the average consumer. If they are able to upsell 3 core Phenoms and cheaper X4s to the average Joe, they should hold out alright.

Their strategy should be to start bundling Phenoms + Mobo and/or 5xxx videocards anywhere and everywhere.

 

[Idontcare]

Originally posted by: KingstonU
The current spin is (RB-C2? CPU-Z) has helped Phenom a lot but has been in production for how long now? It seems to have a thermal wall since the 965 is rated at 140W TDP. Is there at least new-spin of Deneb scheduled to take it 3.6-4.0+ GHz at stock and lift the overclocking ceiling of 4.0GHz to something like 4.5+GHz?

The only other things on the horizon are desktop version of the 6-core Istanbul and Opteron's 12-core Magny-Cours. Then I guess HD6000 series GPU's by 2H 2010.

There is a C3 waiting in the wings, rumored to do for PhII what the G0 stepping did for the Q6600. Supposedly enables at 140W 3GHz Magny-Cours, cut that in half for a Thuban and watch those clocks climb as you eat up the rest of your 70W budget.

Some places have the C3 "listed" but not shipping, its supposed to be under NDA, last I read was NDA lifts Nov 4. (I think that was the date IIRC)

But yeah, until 2011, there was a cc with AMD about 2 weeks ago where Dirk (or some other authority on the matter in AMD) said BD is a 2H 2011 deal now, but that Fusion (Deneb shrink quadcore + GPU monolithic) would be first on 32nm and would be Q4 2010 for sure.

 

[Eeqmcsq]

AMD is supposed to have a new C3 stepping coming, plus the 6 core Phenom II code named Thuban. Otherwise, it will be mostly higher binned CPUs, higher clocks, etc to keep them afloat in the low to mid end markets.

AMD has a wide range of 45w CPUs in the Athlon IIs. I'd like to see them get OEMs into making small form factor PCs, nettops, and all-in-one PCs. And speaking of 45w, I'd like to see if the new stepping can drop TDPs further, like down to 35w or even 25w. How low can AMD go???

 

[aigomorla]

Originally posted by: Eeqmcsq
And speaking of 45w, I'd like to see if the new stepping can drop TDPs further, like down to 35w or even 25w. How low can AMD go???

That would be very very interesting...

Im still waiting for the ultimate chip to throw on a miniPC for my car.
Atom's too slow, and LGA775 draws too much power still, for my liking.

Need that perfect ratio of power draw + efficiency.

So all the high end processors im used to are automatically thrown out the door.

 

[v8envy]

aigo, have you tried disabling all but 1 core and HT on an i7 and volting it down to .8v or so at 100 mhz base clock * 12 multiplier with 800 mhz RAM? Should still beat the pants off an Atom and sip power. I know at .87v and 1.6 ghz my i7 didn't even break 55C on a stock heatsink running prime95 x 8.

 

[cbn]

How long till AMD makes a 32nm quad core?

 

[CurseTheSky]

Originally posted by: aigomorla

Originally posted by: Eeqmcsq
And speaking of 45w, I'd like to see if the new stepping can drop TDPs further, like down to 35w or even 25w. How low can AMD go???

That would be very very interesting...

Im still waiting for the ultimate chip to throw on a miniPC for my car.
Atom's too slow, and LGA775 draws too much power still, for my liking.

Need that perfect ratio of power draw + efficiency.

So all the high end processors im used to are automatically thrown out the door.

I'm not sure if you can buy them separate, but the new CULV processors coming in some notebooks (like the ASUS UL30A-A2 I just ordered) should fit the bill. Core 2 Solo SU3500 (single core) with a 5.5w TDP, or Core 2 Duo SU7300 (dual core) with a 10w TDP. AMD has a similar line... Athlon LE or something like that.

The SU7300 is essentially just a 1.3 GHz Core 2 Duo, but it's guaranteed to work at low voltages. I heard some reports of half-way decent overclocks on them as well. If 1.8-2.0GHz was possible at 15-20 TDP or so, it would be pretty a impressive chip.

 

[Idontcare]

Originally posted by: Just learning
How long till AMD makes a 32nm quad core?

Q4 2010 at absolute earliest.

 

[KingstonU]

Originally posted by: Idontcare

Originally posted by: Just learning
How long till AMD makes a 32nm quad core?

Q4 2010 at absolute earliest.

And it's still up the air as to if this is going to be a AM3 supported 32nm Phenom (PhIII?) correct? I can't imagine them just skipping the procedure of "shrinking current architecture (Deneb) before introducing a new architecture (BD)". Moving onto a new process node and new architecture at the same time has had catastrophic results for CPU and GPU makers.

 

[Idontcare]

Originally posted by: KingstonU

Originally posted by: Idontcare

Originally posted by: Just learning
How long till AMD makes a 32nm quad core?

Q4 2010 at absolute earliest.

And it's still up the air as to if this is going to be a AM3 supported 32nm Phenom (PhIII?) correct? I can't imagine them just skipping the procedure of "shrinking current architecture (Deneb) before introducing a new architecture (BD)". Moving onto a new process node and new architecture at the same time has had catastrophic results for CPU and GPU makers.

The only comments we have regarding the 32nm deneb shrink are from that most recent conference call in which they (Dirk? I can't remember exactly who it was on the AMD side that made the comment) said it would their first monolithic Fusion chip.

So when it comes to socket questions the question we are really asking is will Fusion-type products have access to all the necessary I/O if they are shoe-horned into an AM3 socket? Or will Fusion require a new socket to enable the GPU stuff to get off the chip as well?

We also don't know if AMD will recycle the 32nm deneb shrink core logic area and create a non-IGP non-fusion dedicated CPU. We can fathom them doing this considering what they have done with Athlon II X2 masksets on 45nm but they have not given any roadmap indications that this will come to pass on 32nm.

All we have is confirmation that 32nm deneb shrink with integrated graphics (fusion) is coming Q4 2010 and BD won't be out until 2H 2011.

 

[Soleron]

@Idontcare

You're right in that Fusion will be monolithic and use a Deneb-class core. That much is officially confirmed.

But nowhere have they said Llano is Q4 '10. It's nowhere in the conference call. 32nm will ramp in late 2010, but you can't guess the retail parts schedule from that, I believe AMD's 45nm was ramping in Q3 '08. Only server Bulldozer parts are 2H '11 confirmed, that says nothing about desktop parts.

I'm skeptical about a die-shrunk Phenom (not Llano) in between now and Bulldozer. It doesn't make economic sense.

I don't think AMD is in a bad position - they are profitable now, and in the sub-$200 market Intel isn't doing anything new until Sandy Bridge - they're sticking with Lynnfield until at least Q4 according to leaked roadmaps. Clarkdale may be 32nm but AMD will be positioning tri- and quad-cores against it (Athlon II X3 and X4 are really low in price).

 

[edplayer]

Originally posted by: KingstonU
What will AMD do between now and bulldozer?

Lower prices

a lot

 

[Idontcare]

Originally posted by: Soleron
But nowhere have they said Llano is Q4 '10. It's nowhere in the conference call. 32nm will ramp in late 2010, but you can't guess the retail parts schedule from that, I believe AMD's 45nm was ramping in Q3 '08.

Well GF pretty much outed them regarding the earliest AMD will have tape-outs for 32nm, so we know regardless what they have planned internally for 2010 they have communicated to GF's to not count on them having anything taped out for 32nm until Q3 2010.

Since it was to GF's own detriment (negative press) to publicly state 32nm production would not happen before Q3 2010 I think we can have confidence in this.

Originally posted by: Soleron
Only server Bulldozer parts are 2H '11 confirmed, that says nothing about desktop parts.

And for AMD the server parts precede the desktop parts which precedes the mobile parts.

I can't tell from your wording, but are you basically saying bulldozer for desktop in 2011 is unlikely gicen that Shanghai for desktop (i.e. Deneb) was out of the picture for 2008? (ditto the istanbul->thuban transition as well?)

Originally posted by: Soleron
I'm skeptical about a die-shrunk Phenom (not Llano) in between now and Bulldozer. It doesn't make economic sense.

I'll admit I'm not following here, how does it not make economic sense? Presumably on 32nm a quadcore (or six-core) Deneb shrink would have higher clockspeed potential at the same TDP...why would they not want that to boost their gross margins?

Originally posted by: SoleronI don't think AMD is in a bad position - they are profitable now

That's what I deplore about this "AMD the Product Company" versus "AMD the company" reporting...it gives rise to this misperception that AMD the product company is doing great.

What they did was setup the accounting such that AMD the Product Company was undercharged for the foundry costs of fabbing those CPU's at GF, as such GF's lost more money than they otherwise would have had they charged AMD the Product Company more for the wafers.

Its a scam. GF could technically charge AMD the Product Company nothing for the wafers, and on top of that they could rebate AMD the Product Company back something silly like $5B and guess what AMD the Product Company would report? Record profits! Except AMD, the business entity that legally exists and must report EPS to their shareholders in accordance with GAAP, would not be able to book this internal-customer sham of a rebate transfer.

Is AMD making profit? No. Were GlobalFoundries to actually charge AMD enough per wafer so that GlobalFoundries were to break-even and stop losing money then you'd suddenly see AMD the Product Company losing more money.

Don't let the smoke and mirrors fool you, when both Peter and Paul work for the same company robbing Peter to pay Paul doesn't mean the company is any wealthier or any less poor.

 

[sonoran]

Originally posted by: Idontcare
Were GlobalFoundries to actually charge AMD enough per wafer so that GlobalFoundries were to break-even and stop losing money then you'd suddenly see AMD the Product Company losing more money.

And sooner or later that will have to happen. AMD is hoping that GF can attract enough other business to take up the slack somewhere along the way. But GF's other customers aren't going to (knowingly) pay a premium to subsidize AMD.

What the spinoff DOES do is allow GF to spread process development costs across all their customers, taking some of that burden off AMD. A wise (and I believe inevitable) move in that respect.

 

[Soleron]

Originally posted by: Idontcare

And for AMD the server parts precede the desktop parts which precedes the mobile parts.

I can't tell from your wording, but are you basically saying bulldozer for desktop in 2011 is unlikely gicen that Shanghai for desktop (i.e. Deneb) was out of the picture for 2008? (ditto the istanbul->thuban transition as well?)

AMD's John Fruehe says that isn't neccessarily how it will be this time - the server and desktop teams are completely seperate now (they don't even share dies starting with Istanbul). So I'm saying it could be earlier. Desktop will be more in need of a refresh than server as Magny-Cours is looking very competitive on price and performance.

I'll admit I'm not following here, how does it not make economic sense? Presumably on 32nm a quadcore (or six-core) Deneb shrink would have higher clockspeed potential at the same TDP...why would they not want that to boost their gross margins?

Many reasons. I advise you look through John Fruehe's posts on AMDzone which give an insight into why AMD doesn't do certain things.

Firstly, launching 32nm as soon as they have it will mean it could be more expensive to produce than 45nm, because cost falls rapidly from introduction. AMD launches products on the part of the cost vs. time curve that allows them to be profitable.

Second, process node does not determine clock speed or thermals as much as you think. According to JF, architecture is far more important. 90nm Athlon 64s could still win over 65nm Pentium 4s as an example.

Third, the process of doing a shrink is much more expensive than is usually assumed. The design has to be done from scratch with the new node, and if the product only has a 6 month lifecycle as a 32nm Deneb would have I don't think the profit would justify the cost.

Finally, what market would it sell to? They could only launch one derivative (just desktop, and pick one of <$75,$75-$150, and $150+ price ranges) in six months, say the quad-core. Top-to-bottom rollout takes a year, look at past AMD and Intel architectures. Even if the quad-core could clock 10% higher (and if we take Conroe to Penryn as a typical shrink that's true) then it wouldn't actually move anywhere against Intel - Nehalem would still be a better choice above $200 as Intel would just bump the i5 750 and above by one speedbin to compensate, at almost no cost. But the testing and qualification costs would be very high (tens of millions). Finally, no OEM would pick it up with such a short cycle.

That's what I deplore about this "AMD the Product Company" versus "AMD the company" reporting...it gives rise to this misperception that AMD the product company is doing great.

[...]

Its a scam.

I understand your feeling on this, and many companies do do this, but I don't think that's true. GF is severely underutilised (in fab capacity) - this was stressed in the conference call. I think AMD is paying them a reasonable amount, but with all this unused equipment it isn't enough.

But this will change fast. Once they have bulk processes they can do manufacturing for others. They are in talks with many companies (ST Micro being the only public one) and if companies are to use those new nodes they have to start designing for them now and must therefore be in talks with them. GF wouldn't have rearranged their bulk 28nm/32nm/40nm process schedule a few months ago if they weren't responding to such invisible but real demand. When these companies start using the spare capacity of the fabs, GF will be profitable. They will also reduce the share of passed-on capital costs AMD is paying.

Finally, AMD isn't manufacturing GPUs there yet but they clearly will soon (probably 28nm) so that will be a huge increase in utilisation and a transfer of revenue for TSMC to GF.

So I think GF's unprofitability is purely a reflection of their lack of customers and lack of bulk processes at present. If they're still loss making when they have a 28nm process serving two or more non-AMD customers then we can call it unsustainable.

 

[Phynaz]

Dude, do you have any idea at all who you trying to lecture to about chip design and manufacturing?

Go back and read some of IDK's posts before you "advise" him to look at a AMD marketing employee's posts.

 

[aigomorla]

Originally posted by: Phynaz
Dude, do you have any idea at all who you trying to lecture to about chip design and manufacturing?

LOLZ

You made me spit coffee at my monitor...

Yeah IDC is as bright as it gets on chip manufacture.
Theres a reason why he's L33t!

 

[richierich1212]

Originally posted by: Phynaz
Go back and read some of IDK's posts before you "advise" him to look at a AMD marketing employee's posts.

John Freuhe isn't a marketing employee. "John Fruehe is the Director of Business Development for Server/Workstation products at AMD."

But whatever, I just want new chips to play with!! Q4 2010 is still a year away .

 

[RussianSensation]

Originally posted by: Soleron

Firstly, launching 32nm as soon as they have it will mean it could be more expensive to produce than 45nm, because cost falls rapidly from introduction. AMD launches products on the part of the cost vs. time curve that allows them to be profitable.

I thought cost is both a function of yields and volume of chips being manufactured using new technology? The higher the chip yields, the lower the waste per wafer. But aren't yields achieved over time in improvements in continuous manufacturing? Secondly, your capital asset investment cost for new machinery and equipment should continue to fall as you amortize your costs over the larger volume of production. In other words, assuming your yields are decent, you should in fact start production as soon as possible since with time your 'new technologies' are only aging if you aren't manufacturing anything....you wouldn't build a 32nm factory and just let it sit idle. However, investment in new technologies requires money. The reason AMD doesn't have 32nm before Intel is because they have no $$$ to build FABs anymore.

Second, process node does not determine clock speed or thermals as much as you think. According to JF, architecture is far more important. 90nm Athlon 64s could still win over 65nm Pentium 4s as an example.

Dude....that's just crazy talk! You can't have one without the other. Try building a Nehalem on 180nm....

65 - 45 - 32nm improvements
"Current flows from source to drain when the transistor is on, and it isn?t supposed to flow when it?s off. Now as you shrink the transistor, all of its parts shrink. At 65nm Intel found that it couldn?t shrink the gate dielectric any more without leaking too much current through the gate itself. Back then the gate dielectric was 1.2nm thick (about the thickness of 5 atoms), but at 45nm Intel?s switched from a SiO2 gate dielectric to a high-k one using Hafnium. That?s where the high-k comes from.

The gate electrode also got replaced at 45nm with a metal to help increase drive current (more current flows when you want it to). That?s where the metal gate comes from.

The combination of the two changes to the basic transistor gave us Intel?s high-k + metal gate transistors at 45nm, and at 32nm we have the second generation of those improvements.

The high-k gate dielectric gets a little thinner (equivalent to a 0.9nm SiO2 gate, but presumably thicker since it?s Hafnium based, down from 1.0nm at 45nm ) and we?ve still got a metal gate."

A lower node allows you to fit more transistors within a given powerplane/power consumption levels. Transistor resistance decreases with lower nodes. This will allow you to you introduce faster transistors/switching, which will in turn allow you to increase clock speeds. I do realize that clock speed is also a function of how complex the processor pipeline (stages in it) as as well, not just the node. As far as I can see it, one of the major advantages Intel has always had over AMD is being 12-15 months ahead in node shrinkage. Not to mention that with smaller die, there should be less wafer waste.

"We?re looking at the comparison of leakage current vs. drive current for both 32nm NMOS and PMOS transistors. The new transistors showcase a huge improvement in power efficiency. You can either run them faster or run them at the same speed and significantly reduce leakage current by a magnitude of greater than 5 - 10x compared to Intel?s 45nm transistors." <<<<<<<<<<<<!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!>>>>>>>>>>>

If node process wasn't that important, then how is it that R600 (HD2900XT) -> RV870 (HD5870) is providing almost 4x the performance increase at lower power consumption levels via more or less the same architecture?

 

[Soleron]

Well, AMD's technology development is behind Intel's for sure, that's not a technical decision. That's likely because they have less capital to throw at it, and less volume to make up the capital costs.

What I'm trying to say is that AMD could launch a 32nm chip now. It would have sub-10% yields, be leaky, poorly-clocking and all-around inferior to 45nm, and it would cost several times as much per good die to produce. But it could be done. So there must be a point at which it goes from being uneconomic to do a 32nm chip to a good decision. And I think when GF talks about Q3 2010 as "risk production" and "volume ramp" they are referring to a technical position, which is unlikely to be when retail chips are viable. Look at when TSMC announces processes - 40nm was "ready" in late 2008, but the first retail chips in any quantity are in September this year.

I'll make another point against 32nm Deneb, which is that every dollar and man-hour you spend on it is less to spend on Bulldozer. And Bulldozer better be perfect and on time.

...we should know for certain either way on the November Analyst Day.

--

That said, I have no insight on process technology beyond what tech reviewers, news reports, and JF say. If you know that AMD's process technology can launch 32nm Denebs profitably in Q3 '10 then I'd love to know about it.

 

[Idontcare]

Originally posted by: RussianSensation
*snip*

If node process wasn't that important, then how is it that R600 (HD2900XT) -> RV870 (HD5870) is providing almost 4x the performance increase at lower power consumption levels via more or less the same architecture?

You mean Moore's law and node cadence has something to do with performance scaling too? :laugh: And here I thought Moore's law was something about the number of instructions in an ISA doubling every 12-24 months or some such.

(they don't, in case you were wondering)

What if you built a Core i7 using 1000nm transistors?

Intel has been on an integration rampage lately. Bloomfield integrated the memory controller and Lynnfield brought the PCIe controller on-die. Sean held up an example of what would happen if Intel had stopped reducing transistor size back in the 386 days.

Here's an image of what the Core i7 die would look like built using ~1000nm transistors instead of 45nm:
http://images.anandtech.com/re...te%202/hugenehalem.jpg

Assuming it could actually be built, the power consumption would be over 1000W with clock speeds at less than 100MHz.

So why do we have 4GHz 150W Nehalems instead of 0.1GHz 1kW Nehalems? ARCHITECTURE! No, wait...what?

The truth is architecture/design/layout and process technology all play equal roles in enabling cost-effective performance, it is a red flag if someone tries to talk as if one dominates all others.

 

[aigomorla]

Originally posted by: Idontcare

So why do we have 4GHz 150W Nehalems instead of 0.1GHz 1kW Nehalems?

i thought the .1ghz 1kw nehalems were called itaniums.

 

[Kuzi]

Originally posted by: Soleron

I'll admit I'm not following here, how does it not make economic sense? Presumably on 32nm a quadcore (or six-core) Deneb shrink would have higher clockspeed potential at the same TDP...why would they not want that to boost their gross margins?

Many reasons. I advise you look through John Fruehe's posts on AMDzone which give an insight into why AMD doesn't do certain things.

Firstly, launching 32nm as soon as they have it will mean it could be more expensive to produce than 45nm, because cost falls rapidly from introduction. AMD launches products on the part of the cost vs. time curve that allows them to be profitable.

Actually many companies do just that. Release a shrunk down version of an old architecture on a new process, thus allowing them to master the new process, improve yields, and produce larger chips in the future.

Second, process node does not determine clock speed or thermals as much as you think. According to JF, architecture is far more important. 90nm Athlon 64s could still win over 65nm Pentium 4s as an example.

True, but clock speed and thermals usually always improve when shrinking the same architecture to a smaller process node. Plus you get a much smaller die size.

Third, the process of doing a shrink is much more expensive than is usually assumed. The design has to be done from scratch with the new node, and if the product only has a 6 month lifecycle as a 32nm Deneb would have I don't think the profit would justify the cost.

Why would it have a 6 month life-cycle? Not all people will require an 8-core Bulldozer running at 4GHz. Notebooks, Netbooks, HTPCs, all would benefit from small, power efficient dual and quad core cpus. Also an 8-core version of a 32nm Deneb (2x MCM) would compete well in the server space, as it would end up being smaller than a 6-core Istanbul and probably run at higher clocks.

Finally, what market would it sell to? They could only launch one derivative (just desktop, and pick one of <$75,$75-$150, and $150+ price ranges) in six months, say the quad-core. Top-to-bottom rollout takes a year, look at past AMD and Intel architectures. Even if the quad-core could clock 10% higher (and if we take Conroe to Penryn as a typical shrink that's true) then it wouldn't actually move anywhere against Intel - Nehalem would still be a better choice above $200 as Intel would just bump the i5 750 and above by one speedbin to compensate, at almost no cost. But the testing and qualification costs would be very high (tens of millions). Finally, no OEM would pick it up with such a short cycle.

Again I see many uses for a Phenom III/Athlon III that would justify the cost of the shrink. AMD it seems won't be able to compete with Intel at the high end for a while, so they have to look at other areas to be profitable.