What is the truth about nVidia?

[MTd2]

TSMC's Unified Design For Manufacturing (UDFM) architecture, debuting Monday (June 9), is touted as giving chip makers unprecedented access--at no charge--to its proprietary simulator, DFM models and other production information for 32 nm. The goal is to give IC makers a head start with TSMC's 32-nm process, which is expected to move into production by the end of 2009.

To address those challenges, TSMC will offer a DFM Design Kit (DDK) encapsulating two technologies over a common application programming interface: the foundry provider's proprietary simulator and its DFM data. Customers doing 32-nm designs can download the encrypted DDK from TSMC's portal at no charge.

Nice, that means that nVidia can start projecting its 32nm right now, and start production right at 4Q 2009. And I just can't see why SUN is more worried than nVidia. nVidia portifolio is still weak with CPUs, so it's better to guarantee that their GPGPU chips will be much better than Intel's or AMD's.

And on the contrary nVidia has larger in this kind of approach than Intel and AMD. Look at this article, from 2005. http://www.chipdesignmag.com/display.php?articleId=55

[AussieFX]

Does DFM mean AMD can no longer use APM/CTI (or whatever it's called) where they can make changes midstream?

[wyrmrider]

IN this thread
APM is for AMD manufactured CPU's- including Dresden and Chartered
DFM for TMSC manufactured ATI/AMD GPU's
The important thing is that AMD has extensive experience with several design and manufacturing tools
NVIDA dosen't
AMD will be using DFM for it's 45nm CPU's also

[abinstein]

To address those challenges, TSMC will offer a DFM Design Kit (DDK) encapsulating two technologies over a common application programming interface: the foundry provider's proprietary simulator and its DFM data. Customers doing 32-nm designs can download the encrypted DDK from TSMC's portal at no charge.

Nice, that means that nVidia can start projecting its 32nm right now, and start production right at 4Q 2009. And I just can't see why SUN is more worried than nVidia. nVidia portifolio is still weak with CPUs, so it's better to guarantee that their GPGPU chips will be much better than Intel's or AMD's.

The assumption here is that TSMC is ahead of Intel w.r.t. 32nm DFM. Of course, this is not true.
Nvidia/TSMC has never been able to advance to smaller node before Intel in recent years. What makes you think they can do so at 32nm?

And on the contrary nVidia has larger in this kind of approach than Intel and AMD. Look at this article, from 2005. http://www.chipdesignmag.com/display.php?articleId=55

The article talks about nvidia's design flow. It does not in any way support your claim that nvidia is larger in this kind of approach.

[abinstein]

IN this thread
APM is for AMD manufactured CPU's- including Dresden and Chartered
DFM for TMSC manufactured ATI/AMD GPU's
The important thing is that AMD has extensive experience with several design and manufacturing tools
NVIDA dosen't
AMD will be using DFM for it's 45nm CPU's also

I thought APM and DFM are at two different levels, where APM is at the manufacturing level, done at the fabrication plant, and DFM is at the design level, done by the designers.

What I understood --

APM (Automated Precision Manufacturing) -- a set of programs that detect and analyze chip properties and feed the result back to the manufacturing tools for fine tuning

DFM (Design For Manufacturability) -- rules and tools specific to each manufacturing technology for process simulation at the design time to obtain optimal yield and performance

[MKruer]

I thought APM and DFM are at two different levels, where APM is at the manufacturing level, done at the fabrication plant, and DFM is at the design level, done by the designers.

What I understood --

APM (Automated Precision Manufacturing) -- a set of programs that detect and analyze chip properties and feed the result back to the manufacturing tools for fine tuning

DFM (Design For Manufacturability) -- rules and tools specific to each manufacturing technology for process simulation at the design time to obtain optimal yield and performance

I have always though of it in this way. DFM is to

Code: Select all

Original Design => Durring FAB process becomes
   +-----             /----
   |                 /
   |  +--     =>     |  /--
   |  |              |  |

DFM Design => Durring FAB process becomes

+-+
| +---              +-----
++                  |
  | +--      =>     |  +--
  | |               |  |

This is an over all design modification to the lithography design. APM is different in that its a case by case improvement. Perhaps the best way to describe is that not all waffers are the same, and that some require additional processing during the lithography process. bad example but say the first waffer requires requires 5 minutes to etch a layer. the second wafer might require 6 minutes. with out the APM the second wafer would be treated like the first, but the problem is because the second was not fully etched, it could lead to stability issues with the chip. APM is to fix the defects on a wafer by wafer basis.

[scientia]

Nice, that means that nVidia can start projecting its 32nm right now

Well, not until Q3 when it becomes available.

and start production right at 4Q 2009.

Yes, if they begin production without bothering to test first.

And I just can't see why SUN is more worried than nVidia.

Perhaps because Niagara is barely holding on while nVidia is currently #1 in discrete graphics.

And on the contrary nVidia has larger in this kind of approach than Intel and AMD. Look at this article, from 2005. http://www.chipdesignmag.com/display.php?articleId=55

This article has nothing to do with DFM.

[scientia]

As has already been said, APM is the control system that keeps the FAB tooling within spec. The tooling slowly degrades after maintenance so the software dynamically adjusts the production process. The individual tools also have slight differences and again APM tries to automatically adjust for these differences. To some extent APM is also able to adjust for variation to try to improve yield. In other words, a batch that might otherwise have been scrapped might still be saved at a lower clock level. The point however as I've made elsewhere is that design requirements at 65nm have gotten so tight that AMD may actually be selling chips that would all have been scrapped at 130nm and mostly scrapped at 90nm. AMD learned a hard lesson at 65nm which I think has benefitted the ATI designs at 55nm. In contrast, I think nVidia is just now learning the same lesson and this will probably be too late for 40nm.

DFM is of course Design For Manufacturability. Intel successfully used this to decrease the resolution requirements of their chips at 65nm. In other words, by altering the design they made the chips easier to make. This allowed them to avoid phase shift masks. But, the resolution requirements at 45nm are so high that Intel will need phase shift masks along with double patterning. Again, this competely blindsided AMD because it hadn't been an issue at 90nm.

However, I've already talked about Intel's and AMD's uses of DFM on CPU design in the Technical Discussion of Manufacturing thread. Interestingly enough, a search on Google for "Intel 65nm K1 DFM" turns up that thread as the 11th hit.

To understand the difference look in the Technical Disscusion thread at the quote from Yan Borodovsky. DFM was not a minor change for Intel.

[Bjorn]

That was fast. Fudo says Nvidia will drop prices by $90/$30 for the GTX280/260 tomorrow.

http://www.fudzilla.com/index.php?optio ... 2&Itemid=1

[wyrmrider]

Then Fudo points to the other shoe dropping http://www.fudzilla.com/index.php?optio ... &Itemid=34


http://www.marketwatch.com/News/Story/Story.aspx?guid={E4D2B8AE-D0D0-4F2C-AE92-F037209C5811}&siteid=yhoof2

On APM and DFM
others have posted references
My point is that AMD has Experience now with both- Nvida does not ( or at least is well into DFM with latest designs/ shrinks)

[Woofermazing]

Just gets worse...or better for AMD, go puma go!
http://www.theinquirer.net/gb/inquirer/ ... a**-itself

[abinstein]

Maybe I'm just naive but I don't think nVidia's problem is AMD's gain. AMD's #1 enemy is Intel, which anyone with a clear mind knows that it plays leaps and bounds dirtier than nVidia or any company on earth. If nVidia becomes weaker, it will bow lower to Intel's monopoly force, which in the end hurts AMD and the whole industry.

X86 is an instruction that should've been gone long ago but got life-supported by Intel's monopoly tactics. Now Intel's trying to put x86 into graphics? Please guys, if not for x86, with the same engineering effort the industry has put into PC, we could've been running 4-5GHz Power6-like CPUs on our desktops! Lets still hope nVidia and its GPGPU gets enough momentum to stop Larrabee.

[T800-101]

This is incorrect. You are assuming that 4850's X2 performance will be equal to its Crossfire performance. This was true with 3870 because the X2 used the same Crossfire link to connect the two parts. However, this does not apply to 4850 because it uses a proprietary link. This means that the software will see 4850 X2 as a single card with one GPU. Scaling will not be an issue.

I'm a bit surprised that you are saying the X2 will be transparent to software. Don't get me wrong, this would be a huge achievement for AMD to say the least. But I remain a bit skeptical on this.

I'm a bit confused on what exactly the proprietary bus link is used for.

[Woofermazing]

Maybe I'm just naive but I don't think nVidia's problem is AMD's gain. AMD's #1 enemy is Intel, which anyone with a clear mind knows that it plays leaps and bounds dirtier than nVidia or any company on earth. If nVidia becomes weaker, it will bow lower to Intel's monopoly force, which in the end hurts AMD and the whole industry.

Well, I'm extrapolating a bit, but if there is a quality issue mobile GPU's from Nvidia, that could mean there is an issue with any Intel based laptop with graphics capable of going toe to toe with a puma setup. It's probably not that serious, but you never know.

X86 is an instruction that should've been gone long ago but got life-supported by Intel's monopoly tactics. Now Intel's trying to put x86 into graphics? Please guys, if not for x86, with the same engineering effort the industry has put into PC, we could've been running 4-5GHz Power6-like CPUs on our desktops! Lets still hope nVidia and its GPGPU gets enough momentum to stop Larrabee.

My memory is pretty vague, but wasn't Intel planning on having Itanium filter down to the desktop, and then AMD foiled that with the Athlon 64?

[mmarq]

The irony is that RV770 is good news for Intel too. More so than some may realize.

Around Nov 2006, nVidia launched 8800GTX followed a bit later by 8800GTS. In the same time frame, nVidia launched the 680i SLI chipset for Intel C2D cpus. These two events were a blow to AMD in the enthusiast space, but these launches were also disasterous for Intel's 975x enthusiast chipset. The delayed launch of Radeon HD 2900XT caused Intel's own Crossfire enabled boards to take a massive sales dive, with the result being that by the beginning of 2007, almost anyone who was using Conroe in a high end system was sticking it in a nVidia board equipped with 1 or 2 nVidia graphics cards. In other words, the only thing Intel was selling was the CPU.

hmmm... no it wasn't... a good news for Intel.

Nvidia is perfectly well controllable by Intel. Nvidia don't have all the requisites to maneuver in the most important of all *THE PLATFORM*... they have to be content with the left overs. When this multiple GPU board caughted fire Intel had to do something about it,... Intel started a FUD war with Nvidia because they (NV) were getting too greedy... but in the end it is nothing but a tempest in a tea cup... even more because Intel will have its Larabee GPGPU on a closed platform i suspect...

Remember the recent USB 3 "war" !?... IF or WHEN Intel and AMD close down their platforms, Nvidia is dead. period... and no matter what, if they can have the best GP/GPU or whatever(which clearly they don't have right now), but they are as good as dead...

AMD is an entirely different story.

Had AMD beted on HT4.0 or 5.0... with this success of RV770, foreseable to clearly be transportable until the end of 2009, simply because the RV770 at TSMC 40nm will be a smaller than 200mm², and better, it could have several clock domains (like NV) and have its SPUs (800 of them) at the double of speed of the rest of the chip... and have a relative small chip (<200mm² ) with ~2,5 TERAFLOPS... simply amazing...

But as i was saying this amazing GP/GPU with a superior platform with a superior HT ( still can't imagine why AMD dumped >70Millions on Rambus lap and can't get their interconnect to have at least a 8x data rate) and with IOMMU... 'will??!'... or 'would' have meaned a tremendous kick on the groin to Intel... even with an inferior CPU!...

Not good news to Intel at all... even by the other point of view!...

With AMD chicken about their platforms, following the PCIe crap and probably bowing to the master and dumping yet more millions for licensing CSI/QPI when they have better,... and with only a REAL competitive CPU when at 45nm HKMG... if it wasn't for this ATI success Intel would had killed AMD in a year.... so they must be in grief because they saw a chance to be the absolute and undisputed dominate force in all of IT hardware industry, and without the propaganda that flows by the ton, because by the end of 2009 AMD would had kaputed...

Now that chance is finished for the time being... so not good news for Intel i'm sure... specially for that criminal organization top "godfather "innis"" bosses, i thing they feel not good news...

And on the other way around, AMD knows the real enemy is Intel, and that if they help keep Nvidia alive they will be better. So the more reason for not following PCIE or CSI/QPI, and betting on an advanced HT with Nvidia ONLY... because if those (PCIE;CSI/QPI) are the most dominant and pervasive and so AMD follows and HT get losted,... then at any other bad step by AMD, Intel can advance for a kill very easily... very easily indeed !... (that is if we can count that AMD is off the hook by now!?)

[abinstein]

My memory is pretty vague, but wasn't Intel planning on having Itanium filter down to the desktop, and then AMD foiled that with the Athlon 64?

I believe AMD supported x86 for a good reason: they got a very good implementation (at that time) of the ISA, K7, which runs faster and scales better than any other implementation including Intel's P6 at that time. With any other instruction set (Power, MISP, EPIC, ...) AMD would've been non-competitive at all.

When it became pretty clear that Intel's improved P6 is no inferior than K8 or K10, AMD changed to say that "ISA doesn't matter - let x86 stay and be everywhere" and focused on scalability (native quad-core) instead. This is not a bad strategy for the company, because it still makes nothing but one type of CPU well: x86 compatibles. However, now AMD has its own GPU microarchitecture, totally different from x86, and is looking at fusion. When Intel makes x86 the ISA of choice for an universal accelerator replacing the role of GPU, will AMD still endorse "x86 everywhere" as it used to?

[AndyW39]

The CPU side is no longer treading water. At 2.5Ghz with quad core AMD has finally hit the mainstream segment.

Thanks for all your points.

I tend to disagree with you on this point though, at least on the desktop side. AMD prices are more value side than mainstream side, even if the performance it getting better. Performance per watt is still poor compared to Intel 45nm. I don't see AMD making headway until 45nm and the execution of this has to be good with now screwups this time. Early indications of speed grades is promising though with speeds slightly more than the 65nm chips.

[mmarq]

When it became pretty clear that Intel's improved P6 is no inferior than K8 or K10, AMD changed to say that "ISA doesn't matter - let x86 stay and be everywhere" and focused on scalability (native quad-core) instead. This is not a bad strategy for the company, because it still makes nothing but one type of CPU well: x86 compatibles. However, now AMD has its own GPU microarchitecture, totally different from x86, and is looking at fusion. When Intel makes x86 the ISA of choice for an universal accelerator replacing the role of GPU, will AMD still endorse "x86 everywhere" as it used to?

If AMD has the boldness of sticking for a Fusion Phase 3, then i believe the CPU will jump inside the GPGPU.... SSE6! will follow and those kind of instructions will prevail...

And what are those instructions ?... CISC ? .... RISC ?... VLIW ?

x86 will be done mostly by a plethora of tech at the APU fron-end: "codemorphing"... macro-fusion... fast-path decode ?;... but i believe a new vector (vec2-5 ?) type of instruction will prevail.

Well that is my guess. But as i read in a very good article( i'm looking for the link) if the power or cost of moving an instruction is ~1000 times more than that of making a simple ALU operation, then something like an ultra-threading mechanism as in actual GPGPUs, and lots of ALU clustered together, with multiple caches, and better pre-fetches and faster memory techniques to make the data as close as possible to those ALUs... will be the designs of the future... and so the ISAs will be mostly vectorized, and all programs deeply threaded...

But this will be clearly after 2012. So again contrary to the BS propaganda flow, it will be ATI again to save AMD... and it may came a time when x86 is discarded for good everywhere except at Intel, which dumbly makes x86 their GPGPU basic ISA choice... another Intel BS trick replaced by AVX in reality, or then this x86 mania on graphics can very well represent a severe blow to them... at last, thank god, the poisoner tastes its own venom!

EDIT: here are the links :
http://acmqueue.com/modules.php?name=Co ... 348&page=2

and specially

http://acmqueue.com/modules.php?name=Co ... ge&pid=531
http://acmqueue.com/modules.php?name=Co ... ge&pid=533

the all Mag issue is a good reading on the subject
http://acmqueue.com/modules.php?name=Co ... ssue_id=48

[AussieFX]

Maybe I'm just naive but I don't think nVidia's problem is AMD's gain. AMD's #1 enemy is Intel, which anyone with a clear mind knows that it plays leaps and bounds dirtier than nVidia or any company on earth. If nVidia becomes weaker, it will bow lower to Intel's monopoly force, which in the end hurts AMD and the whole industry.

X86 is an instruction that should've been gone long ago but got life-supported by Intel's monopoly tactics. Now Intel's trying to put x86 into graphics? Please guys, if not for x86, with the same engineering effort the industry has put into PC, we could've been running 4-5GHz Power6-like CPUs on our desktops! Lets still hope nVidia and its GPGPU gets enough momentum to stop Larrabee.

Do Nvidia want to stop Larrabee or join them?
This is a quote from an article I read in Australian PC Authority magazine today.

"Nvidias view is that the next generation of high quality rendering will mix API's and programming with CUDA and other C/C++ languages using rasterisation and ray tracing. Nvidia is putting a lot of money into ray tracing - in paticular, it acquired University of Utah spinoff RayScale as part of its plans.......
Nvidia aim to enable real time rendering and ray tracing in their next generation processors."

I'm not really across what Larrabee does, but to me it reads like Nvidia want to combine todays methods in combination with ray tracing.
Is a stream processor able to ray trace on it's own or is that the domain of x86 cpu's?

[maduroutmb]

My take on all of this is that Intel still wants to buy nVidia. I think that there are credible sources claiming that Intel tried in mid 2007 but was turned down. This would make sense, as Intel has no high-performance stream processor design. If AMD is able to effectively move an APU into its processors, that would have a huge effect on high performance server CPUs running in environments that can take advantage of a stream proceesor APU and mobile processors.

My thinking on Larrabee is that its goal is not so much high performance as competition where it hurts. ATI has already done well by focusing on a lower market segment and building up (remember that even the 48x0 has two R770 cores per chip). Even if Intel can't get that kind of scalability, they can still compete at the low end, which also happens to be much higher volume than the very top. Intel can afford to produce a graphics chip that is subsidized by its processor division, so Larrabee needs to make money only insofar as regulatory commissions are concerned. So, if Intel can mass produce a GPU that does well at the low to mid range and that doesn't need to be a cash cow, they can severely affect nVidia's margins (nVidia is already painted into a corner on this one). That would drive nVidia's stock price down and ultimately force them to agree to a buyout. At that point, Intel gets its platform to compete with AMD at a reduced price, albeit 2 years late.

[MKruer]

The CPU side is no longer treading water. At 2.5Ghz with quad core AMD has finally hit the mainstream segment.

Thanks for all your points.

I tend to disagree with you on this point though, at least on the desktop side. AMD prices are more value side than mainstream side, even if the performance it getting better. Performance per watt is still poor compared to Intel 45nm. I don't see AMD making headway until 45nm and the execution of this has to be good with now screwups this time. Early indications of speed grades is promising though with speeds slightly more than the 65nm chips.

The performance per watt is questionable at best. First off on a chip by chip bases yes Intel has better per watt rating, but its also doing less. AMD chip include part of the Northbridge (i.e. Memory controller, logic, etc...) that Intel also requires to run their system Problem is this. AMDs 2.6 Barcelona is rated at 140 watts. The rest of the system is say only 25 watts for a total of about 165watts MAX. Intel 2.6 is rated at 95 watts, but the rest of the system pulls around 70 watts, for a maximum of 165 watts. Yes you can make the claim that the CPU is more power efficient, but the overall platform may not be. These numbers i pulled out of my head, but depending on the review the over all performance per watt is very competitive. you as the read just need to be aware that there is more involved with running a system then just the CPU, and right now AMD tends to have a better overall platform until you get above their max, but then again they don't have anything that can compete one on one in that area now anyway. On a side not I would like to add how a lot of reviews are very schizophrenic when it comes to power. First they start off with ooh look at how power efficiency something is then proceed to add two high end graphs cards totaling over 500 watts. Why the hell do you care about power efficiency if you are going to blow your power budget on graphics anyway?

[MKruer]

I'm not really across what Larrabee does, but to me it reads like Nvidia want to combine todays methods in combination with ray tracing.
Is a stream processor able to ray trace on it's own or is that the domain of x86 cpu's?

Yes, you can do ray trace on the current video cards. The only trick is that you need a very powerful system to do it. If I remember correctly about 6 months ago they got someone to do ray tracing on the Quake 3 and 4 engine, but the performance was like 30fps @ 640x480 i.e. not that great compared to the current method. If I also remember correctly doing it on the faster quad core was bout 1-3 fps at the same resolution. my best guess is that the video cards are more adept at doing the ray tracing, but that the CPU has the current speed (mhz) advantage

[MKruer]

My take on all of this is that Intel still wants to buy nVidia. I think that there are credible sources claiming that Intel tried in mid 2007 but was turned down. This would make sense, as Intel has no high-performance stream processor design. If AMD is able to effectively move an APU into its processors, that would have a huge effect on high performance server CPUs running in environments that can take advantage of a stream proceesor APU and mobile processors.

My thinking on Larrabee is that its goal is not so much high performance as competition where it hurts. ATI has already done well by focusing on a lower market segment and building up (remember that even the 48x0 has two R770 cores per chip). Even if Intel can't get that kind of scalability, they can still compete at the low end, which also happens to be much higher volume than the very top. Intel can afford to produce a graphics chip that is subsidized by its processor division, so Larrabee needs to make money only insofar as regulatory commissions are concerned. So, if Intel can mass produce a GPU that does well at the low to mid range and that doesn't need to be a cash cow, they can severely affect nVidia's margins (nVidia is already painted into a corner on this one). That would drive nVidia's stock price down and ultimately force them to agree to a buyout. At that point, Intel gets its platform to compete with AMD at a reduced price, albeit 2 years late.

Thats an interesting idea. So intels push in to graphics with Larrabee is the same sort of push that Intel made into RICS with the Itanimum. That is a very risky strategy. Itanimum may have failed miserably but it did succeed in getting rid of other RICS designs just by the threat. The problem is that Nvidia would be like IBM or SUN to intel in the graphics market. Pushing the x86 might backfire big time. Intel might be sitting in the position of having a working GPU that no one wants to use because of lack of support.

[mmarq]

Thats an interesting idea. So intels push in to graphics with Larrabee is the same sort of push that Intel made into RICS with the Itanimum. That is a very risky strategy. Itanimum may have failed miserably but it did succeed in getting rid of other RICS designs just by the threat. The problem is that Nvidia would be like IBM or SUN to intel in the graphics market. Pushing the x86 might backfire big time. Intel might be sitting in the position of having a working GPU that no one wants to use because of lack of support.

I agree with the general view, but...

lack of support... hmmm... i believe that neither at the platform side, because Intel even controls (for AMD shame) parts of other competitors platforms, they have the most effective force in this and can easily pull out Larabee on QPI....

... and on software side the story is the same, even because a x86 base GPGPU would make at least programming more familiar.

The problem is efficiency of the all set, and AMD Fusion push, that could very well make a prevailing ISA set out of the x86 arena. What we are assisting today is the most tremendous push in order to kill AMD.. ever. If AMD gets kaputed, we will see x86 everywhere " ad infinitum and ad obnoxious"... Nvidia simply cant counter alone.

If AMD succeeds we will see SSE5/6/7 vector INT instructions getting prevalence over x86 and Larabee getting to the scrap. It will be worst than Itanic. But it will be a long war, because Intel will get support no matter what... a little like what happened with Itanic.

[MKruer]

I should have said, not lack of support, but lack of people willing to develop for the new design. The problem with this is I don't think that is accurate as well, reason being is because any high level coding language like CUDA or Book+ is very similar to a striped down version of C++. So those that say Larrabee will be easier to program for are full of it. I highly doubt that the programming of Larrabee will be all that different from the programming of CUDA or Book+ What will an issue will go down to performance.