4GHz Deneb OC reported

[bryanW1995]

Originally posted by: Idontcare

Originally posted by: Ocguy31
As far as voltage tolerances for 45nm CPUs, is there a possiblity that the AMD 45nm could withstand more than the 1.4V that is a commonly accepted Max for the current Intel 45nm? Or that Intel's next gen 45nm could withstand more? Or are they made from the same materials, with the same amount of electromigration, etc?

Likely comparable. It depends on soooo many factors but the likely answer is that the chip-to-chip variability in the peak viable voltage is probably much larger than the difference in the mean value for both companies.

Meaning 1.4V isn't likely to kill a chip from either supplier but 1.8V is, how your specific chip handles 1.6V depends mostly on your specific chip and you can't expect to be able to compare your experience with your neighbor's experience in the gray area.

Of more relevance at this point would be the power consumption. 4GHz at 1.5V on Deneb may consume so much power that cooling the thing and remaining small FFT stable just isn't possible without 90C temps.

with amd's 45nm low-k and using exotic materials (germanium?) it might also be possible for it to handle higher voltages than penryns.

 

[Acanthus]

Originally posted by: bryanW1995

Originally posted by: Idontcare

Originally posted by: Ocguy31
As far as voltage tolerances for 45nm CPUs, is there a possiblity that the AMD 45nm could withstand more than the 1.4V that is a commonly accepted Max for the current Intel 45nm? Or that Intel's next gen 45nm could withstand more? Or are they made from the same materials, with the same amount of electromigration, etc?

Likely comparable. It depends on soooo many factors but the likely answer is that the chip-to-chip variability in the peak viable voltage is probably much larger than the difference in the mean value for both companies.

Meaning 1.4V isn't likely to kill a chip from either supplier but 1.8V is, how your specific chip handles 1.6V depends mostly on your specific chip and you can't expect to be able to compare your experience with your neighbor's experience in the gray area.

Of more relevance at this point would be the power consumption. 4GHz at 1.5V on Deneb may consume so much power that cooling the thing and remaining small FFT stable just isn't possible without 90C temps.

with amd's 45nm low-k and using exotic materials (germanium?) it might also be possible for it to handle higher voltages than penryns.

Ive heard nothing of AMD using SiGe

 

[MarcVenice]

I'm not to sure about this, but perhaps AMD overdrive ( the app that is running ) is loading the cores at 100% ?

One thing I never understood is this. If a q6600 can run at 3.0ghz without any extra vcore, then why doesn't Intel release it with higher stock clock speeds, like 2.6-2.7ghz? And this especially true for AMD, they NEED more performance, if some of their CPU's can run at 3.5ghz they can bin some that run at 2.8-3.0ghz, so they can compete with intel?

 

[Idontcare]

Originally posted by: MarcVenice
One thing I never understood is this. If a q6600 can run at 3.0ghz without any extra vcore, then why doesn't Intel release it with higher stock clock speeds, like 2.6-2.7ghz? And this especially true for AMD, they NEED more performance, if some of their CPU's can run at 3.5ghz they can bin some that run at 2.8-3.0ghz, so they can compete with intel?

For Intel it 100% has everything to do with market segmentation and ASP protection. If the 2.40GHz Q6600 can compete with AMD's upper-end quad-core for performance (and best it for power consumption) then there is no benefit to Intel or Intel's shareholders in releasing a 3GHz chip to replace the Q6600.

For AMD the primary reason you don't see them bin-out 3.5GHz chips for sale as 3GHz SKU's is that the TDP for those chips at 3GHz would be fodder for Intel's marketing machine.

Sure there may be 1,000 to 10,000 people around the planet who are geared up to cool down a 250W processor, and those folks will take a 140W TDP and overclock the bejesus out of it and post their 3.5GHz overclock on the web but that doesn't mean AMD can put them on the market and expect 100,000 people to do well attempting the same.

This goes for Intel as well, which is why you don't see 200W 4GHz yorkies on the market.

Originally posted by: bryanW1995
with amd's 45nm low-k and using exotic materials (germanium?) it might also be possible for it to handle higher voltages than penryns.

While AMD does use SiGe for stress management in the channel, the materials of choice for AMD are no more exotic than Intel's choice of materials...and certainly the argument could be made that Intel's choice is more exotic (HK/MG).

edit start: adding links for proof of SiGe use (thanks to poster inf64 at aceshardware)...specifically see the page 6 graph in upper-righthand corner of this AMD document: http://www.amd.com/us-en/asset...ress_Presentation2.pdf

As well as answer 3 (A3) in this AMD Q&A document: http://www.amd.com/us-en/asset...MD__45nm_Press_Q-A.pdf

edit end

The ability to handle voltages depends on many things, but first and foremost it is the electric field (voltage/distance) that defines the problem. So critical dimensions need to be known.

If Intel's M1 pitch is 10% smaller than AMD's then the electric field placed on the M1 dielectric will be 10% higher in the Intel chip if both the Intel and AMD chips are operating at the same voltage.

Likewise if the contact-to-gate distance is smaller for Intel then the electric field on the dielectric material between the contacts and the gate will be placed under higher electric field stress if the same voltage is used on the Intel chip and the AMD chip. Same is true for the gate oxide, and every other insulating material and distance between electrodes thruout the IC.

So the other question is are the dielectic materials the same or can one company's dielectric material of choice (and integration history) be superior to the other's. This we cannot answer without access to their internal reliability and lifetime data.

If both companies were to use exactly the same materials choice and have identical design rules they could still end up with vastly different reliability and lifetimes when operating at the same voltage due to process integration differences that can enhance or deteriorate the material's electrical properties.

This is one of the many reasons it takes 4 yrs to develop a process node for manufacturing. Hitting the design rules takes maybe 1 yr, hitting those dimensions with acceptable time-zero yields takes another 1-2 yrs, making the parts such that they have time-zero yield and can be expected to still function in 10yrs takes another year of development.

 

[Kuzi]

That is great news, can't wait for Deneb's actual release.

I've heard the Deneb ES samples are getting stable OCs around 3.3-3.5GHz with the "new" AMD chipsets, but not 4GHz. So I'm expecting the release samples to do a bit better, a 3.6GHz stable OC would be nice

The problem I see for AMD now, even if they release Deneb soon, Intel will have Nehalem ready, and the 2.66GHz is supposed to sell for $285. Even the fastest release Deneb, probably a 2.8GHz, will be no match for a 2.66GHz Nehalem. Meaning AMD will have to price their top CPU below $280.

Can they make much money doing that?

 

[harpoon84]

Originally posted by: Kuzi
The problem I see for AMD now, even if they release Deneb soon, Intel will have Nehalem ready, and the 2.66GHz is supposed to sell for $285. Even the fastest release Deneb, probably a 2.8GHz, will be no match for a 2.66GHz Nehalem. Meaning AMD will have to price their top CPU below $280.

Can they make much money doing that?

Do they make much money right now? I think that answers your question.

Deneb will be an improvement, sure, but I actually think the performance gap between Intel and AMD will grow with Nehalem, so its kinda back to the drawing board for AMD.

 

[Kuzi]

Originally posted by: harpoon84

Originally posted by: Kuzi
The problem I see for AMD now, even if they release Deneb soon, Intel will have Nehalem ready, and the 2.66GHz is supposed to sell for $285. Even the fastest release Deneb, probably a 2.8GHz, will be no match for a 2.66GHz Nehalem. Meaning AMD will have to price their top CPU below $280.

Can they make much money doing that?

Do they make much money right now? I think that answers your question.

Deneb will be an improvement, sure, but I actually think the performance gap between Intel and AMD will grow with Nehalem, so its kinda back to the drawing board for AMD.

True, but Deneb will be a bit smaller than Phenom in size, so that can help some. It also depends on yields which may not be that good at the start, but will improve over time.

 

[BTRY B 529th FA BN]

When are these Denebs suppose to be released?

 

[Foxery]

Originally posted by: Kuzi
The problem I see for AMD now, even if they release Deneb soon, Intel will have Nehalem ready, and the 2.66GHz is supposed to sell for $285. Even the fastest release Deneb, probably a 2.8GHz, will be no match for a 2.66GHz Nehalem. Meaning AMD will have to price their top CPU below $280.

Can they make much money doing that?

Spin it from a different angle:
Can they loose less money doing that? Yes. The new product:
1) Is physically smaller, thus cheaper to manufacture
2) Cannot possibly be mistaken for one with the TLB bug
3) Should have lower power+heat, which is appealing to enterprise customers (who are already satisfied with AMD performance)
4) Directly competes with Yorkfield's market segment, not Nehalem's. (although if Intel gets too aggressive on pricing... ouch)

It's more about catching up, rather than getting ahead.

 

[Idontcare]

Originally posted by: Kuzi

Originally posted by: harpoon84

Originally posted by: Kuzi
The problem I see for AMD now, even if they release Deneb soon, Intel will have Nehalem ready, and the 2.66GHz is supposed to sell for $285. Even the fastest release Deneb, probably a 2.8GHz, will be no match for a 2.66GHz Nehalem. Meaning AMD will have to price their top CPU below $280.

Can they make much money doing that?

Do they make much money right now? I think that answers your question.

Deneb will be an improvement, sure, but I actually think the performance gap between Intel and AMD will grow with Nehalem, so its kinda back to the drawing board for AMD.

True, but Deneb will be a bit smaller than Phenom in size, so that can help some. It also depends on yields which may not be that good at the start, but will improve over time.

I would argue, given the rather lackluster fab utilization atm, that the overriding cost issue with Phenom (and the concern for carry-over cost issues with Deneb vs. $284 Bloomfield) is the fixed overhead costs of existing asset depreciation combined with capex costs of tooling for 45nm.

Were AMD to get volume of Phenom or Deneb up high enough so as to enter the asset utilization regime considered necessary for sustainability (>80% fab loadings) then the cost-impact of yields and die-size can become a discussion.

But for now, running 1 wafer at 80% yield netting 60 NUB's versus running 2 wafers at 40% yield netting 60 NUB's is hardly an overriding cost issue (it can be when you are running at/near capacity) as much as the fact that one fab is at or near 50% loads while the other is effectively mothballed. (numbers here are merely examples, not claims of AMD's specific yields or NUB's).

This is the dilema of every IDM when possessing excess capacity...yield gains do not produce tangible cost reductions (fabbed wfrs actually cost very little in materials cost, its the fixed assets and headcounts that make up >80% the per wafer costs you see bandied about) nor do improved uptimes and asset management produce increased revenues.

 

[magreen]

Originally posted by: Idontcare
I would argue, given the rather lackluster fab utilization atm, that the overriding cost issue with Phenom (and the concern for carry-over cost issues with Deneb vs. $284 Bloomfield) is the fixed overhead costs of existing asset depreciation combined with capex costs of tooling for 45nm.

Were AMD to get volume of Phenom or Deneb up high enough so as to enter the asset utilization regime considered necessary for sustainability (>80% fab loadings) then the cost-impact of yields and die-size can become a discussion.

But for now, running 1 wafer at 80% yield netting 60 NUB's versus running 2 wafers at 40% yield netting 60 NUB's is hardly an overriding cost issue (it can be when you are running at/near capacity) as much as the fact that one fab is at or near 50% loads while the other is effectively mothballed. (numbers here are merely examples, not claims of AMD's specific yields or NUB's).

This is the dilema of every IDM when possessing excess capacity...yield gains do not produce tangible cost reductions (fabbed wfrs actually cost very little in materials cost, its the fixed assets and headcounts that make up >80% the per wafer costs you see bandied about) nor do improved uptimes and asset management produce increased revenues.

Thank you Idc. I never understood what was so expensive about a wafer of silicon. Now I get it -- the "per wafer cost" is not actually the cost of the wafer. It's more like how many man-hours of work in the Tombstone pizza factory does it take to produce 1000 pizzas, i.e. it's not the cost of the cheese & dough & tomatoes.

 

[Idontcare]

Originally posted by: magreen
Thank you Idc. I never understood what was so expensive about a wafer of silicon. Now I get it -- the "per wafer cost" is not actually the cost of the wafer. It's more like how many man-hours of work in the Tombstone pizza factory does it take to produce 1000 pizzas, i.e. it's not the cost of the cheese & dough & tomatoes.

Precisely. It's the humans doing the making, plus the humans doing the quality control, ingredient purchases, humans doing sales and distributions that create about 30-40% of the cost structure...and the biggie of course is the fricken cost of the factory (building) plus equipment (depreciates). Not too mention our lovely governments and their payroll, medicare, social security, business income, and local sales taxes.

 

[Kuzi]

Originally posted by: BTRY B 529th FA BN
When are these Denebs suppose to be released?

AMD supposedly has "retail" Deneb (C2) stepping going through testing, so if all goes well, it could be released in a few months, around Oct/Nov timeframe.

I hope AMD surprises us and Deneb performs better than expected, +15% over a similarly clocked Phenom.

@IDC, thnx for the informative post, now I understand yields are only a small part of the total equation

 

[Idontcare]

Originally posted by: Kuzi

Originally posted by: BTRY B 529th FA BN
@IDC, thnx for the informative post, now I understand yields are only a small part of the total equation

Please understand though that this is only true when operating with excess existing fab capacity (i.e. supply exceeds demand).

I don't want to mislead anyone. Yields matter very much to revenue and costs when the supplier is supply constrained (as AMD was when they took on the DELL account). In that environment a 1% increase in yields can lead to a 1% increase in revenue which is awesome.

 

[BTRY B 529th FA BN]

If all goes well in the land of competition, we'll see more AMD talk back on the forums

 

[Kuzi]

Originally posted by: Idontcare

Originally posted by: Kuzi

Originally posted by: BTRY B 529th FA BN
@IDC, thnx for the informative post, now I understand yields are only a small part of the total equation

Please understand though that this is only true when operating with excess existing fab capacity (i.e. supply exceeds demand).

I don't want to mislead anyone. Yields matter very much to revenue and costs when the supplier is supply constrained (as AMD was when they took on the DELL account). In that environment a 1% increase in yields can lead to a 1% increase in revenue which is awesome.

I got what you meant IDC, you where talking about the situation AMD is in right now. Yields can matter a lot for "fabless" companies too such as Nvidia.

 

[MarkLuvsCS]

Why do the numbers on Core1-Core3 show up slightly larger and bolder? You can clearly see how their character sizes fill the bars on Core1-Core3, but on Core0 they appear to be fit properly. I'm not saying its BS, but maybe just a programming glitch? I'm more curious than anything. I look forward to healthier competition for cheaper parts for consumers.

 

[dmens]

Originally posted by: Idontcare

Originally posted by: Lonyo
Is stock voltage likely to be any indicator of what sort of upper voltage levels they will be able to take?
I mean, Yorkie seem to be 1.05~1.15 stock vcore, and the screenshot of the 2.3 shows 1.224v.
Any chance that outside the chips just needing a higher vcore they will also be able to survive a higher vcore than the 45nm Yorkfields?

It is directly related by way of something called a shmoo plot in the technical lingo.

A shmoo plot is a plot of stable operating voltage versus clockspeed. Most shmoo plots are rate limiting (initially you observe a rapid rise in peak stable clockspeeds for small increases in voltage, followed later by a period of diminishing returns in which significant voltage increases are required with only small gains in peak clockspeed).

So the farther "to the left" your chip's voltage is at the highest clockspeed then that does speak to the likelihood of your chip not reaching that point of diminishing returns as you ramp up the shmoo curve.

I think this linked example of how shmoo plots can be generated in practice is pretty cool: http://www2.imm.dtu.dk/courses/02208/texts/shmoo2.htm

Here's a real-world shmoo plot (of sorts, you have to flip the x and y axes) I created with my QX6700: http://i272.photobucket.com/al...ableVcoreversusGHz.jpg

The caveat of course is that every CPU has its own shmoo plot. Sure they all should be more or less the same but that doesn't guarantee you don't have an outlier whose shmoo plot just goes dead flat really quick at low GHz.

a frequency/voltage 2d shmoo will not give any information regarding chip lifetime reduction due to overvoltage. the only information given is the expected operating frequency at the testing temperature.

couple basic reasons for frequency falloff at high voltage, thermal constraint and wire delay dominated paths. dunno about deneb but you'll find out soon enough how nehalem behaves.

 

[Nemesis 1]

http://translate.google.com/tr...ie=UTF8&sl=zh-CN&tl=en

 

[3Ball]

Thats an interesting find. I would sure like to play with one. lol

Best,

3Ball

 

[ultra laser]

Those super pi shots show that it's still a lot slower than Intel.

My Conroe at 3.2GHz does 1M in 16 seconds, whereas a 3.2GHz deneb takes 22 seconds.

 

[CTho9305]

Does SuperPI performance correlate with any real-world performance? How well does it predict performance in other applications across architectures? How about between "similar" architectures like banias and penryn? (Honest questions - I don't usually bother remembering synthetic benchmark results)

 

[Zstream]

Originally posted by: CTho9305
Does SuperPI performance correlate with any real-world performance? How well does it predict performance in other applications across architectures? How about between "similar" architectures like banias and penryn? (Honest questions - I don't usually bother remembering synthetic benchmark results)

You can only use SuperPI against the similar architectures in order to determine performance. SuperPI is fit perfect for Intel and rather poor for AMD.

So if you can compare superpi #'s from Deneb and from Phenom you should get the % in overall improvement. That is if Deneb is the exact same arch as the Phenom which is sketchy still.

 

[Gikaseixas]

great OC, things are getting interesting

 

[quadomatic]

Dang...4 ghz on stock? I'm betting it'll get some performance boosts before they release it.

How much will the CPU cost? Hopefully it's lower than the cost of the current Phenom 9650 ($190). I'm probably not ready to upgrade CPU/Mobo/Ram for at least another year or so though. Too bad I'll probably never get an upgrade as cheap as the one I got back in February (E2200+ECS mobo+RAM for $93, oc'd to 3ghz).

Did they make a black edition of the current 9650? I wonder if we'll see black edition of the Deneb.

Hmmm...Phenom 9850 black edition is priced slightly lower than the 9650. What's up with that?