What happens to nvidia?

[Paratus]

Where NV is being out played is in production engineering. For several generations now ATI has been able to cram more xtors/mm than NV. Considering how much cash NV has and how important xtor density is with their big die strategy there is no excuse for the lack of talent in this area.

It also gives AMD an automatic 10-20 percent discount on the number of wafers they buy from TSMC. A discount NV has to make up with whatever sweetheart pricing deal some of you think they have with TSMC.

To make matters worse they completely didn't understand the issues with TSMCs 40nm process - ATIs engineers did and beat them to the market by 9 months.

Plus what I think a lot of folks are missing is how unequally the delay to 32/28nm is affecting each company. NV needs a die shrink to bring their next high end chip. AMD on the other hand can add 1bl xtors, 200nm of silcon, and 75-90watts on top of the 5870 before they equal the GF100. The 5870 that already is about 95 percent of the GTX in Mafia (w/o Physix). They could probably bring the HD6XXX AND HD7XXX series out on 40nm if they had to.

It's going to be a challenging year for NV, they need to get back on their A game.

 

[Scali]

Where NV is being out played is in production engineering. For several generations now ATI has been able to cram more xtors/mm than NV. Considering how much cash NV has and how important xtor density is with their big die strategy there is no excuse for the lack of talent in this area.

Or... because nVidia builds larger circuits, they opt to have slightly lower density in order to increase yields (one reason for lower density could be the use of the venerable double via (or even triple or more) technology!!!!!!!!!11oneone).

To make matters worse they completely didn't understand the issues with TSMCs 40nm process - ATIs engineers did and beat them to the market by 9 months.

This was already debunked in another thread some weeks ago.

AMD on the other hand can add 1bl xtors, 200nm of silcon, and 75-90watts on top of the 5870 before they equal the GF100.

AMD's problem is the other way around:
nVidia has already figured out what to DO with the extra transistors, and has added several technologies to the GF100 architecture that AMD has yet to implement... such as a unified address space for compute shaders (giving full C++ support), ECC support, parallelized triangle setup and tessellation circuitry.

 

[JAG87]

ATi make great hardware, i have never said otherwise, however their drivers suck. NV are risk takers and they have paid the price several times, however they are multi function devices, have a much better driver\software team, are more about customers rather than manufacturers and more times than not leaders in their trade.

/thread

This is what matters to the professional clientele, and that's where the money is.

 

[Paratus]

Or... because nVidia builds larger circuits, they opt to have slightly lower density in order to increase yields (one reason for lower density could be the use of the venerable double via (or even triple or more) technology!!!!!!!!!11oneone).

That doesn't make much sense. AMD doubled up on vias and still manages a better density. Plus why would it be a benefit to make an already large chip more bloated by purposely making it less dense? As I understand it yields go down due to larger dies resulting less dies per wafer, defects per wafer, and poor design not from xtor density. Your point doesn't make a lot of sense to me.

This was already debunked in another thread some weeks ago.

I missed that thread - have a link?

AMD's problem is the other way around:
nVidia has already figured out what to DO with the extra transistors, and has added several technologies to the GF100 architecture that AMD has yet to implement... such as a unified address space for compute shaders (giving full C++ support), ECC support, parallelized triangle setup and tessellation circuitry.

[/QUOTE]


IMHO it's a better problem to have a large space to work in than to already have your back against the wall. Besides GF104 already proves that to be competitive in this space they had to remove some of those features. AMD can add features & performance.

 

[Scali]

That doesn't make much sense. AMD doubled up on vias and still manages a better density.

This was all discussed in detail already.
See this thread:
http://forums.anandtech.com/showthread.php?t=2095030&highlight=double+vias

Plus why would it be a benefit to make an already large chip more bloated by purposely making it less dense?

Uhhhhhh... what did I just say?
because nVidia builds larger circuits, they opt to have slightly lower density in order to increase yields

Benefit underlined.

As I understand it yields go down due to larger dies resulting less dies per wafer, defects per wafer, and poor design not from xtor density. Your point doesn't make a lot of sense to me.

That is because your understanding is lacking.

IMHO it's a better problem to have a large space to work in than to already have your back against the wall. Besides GF104 already proves that to be competitive in this space they had to remove some of those features. AMD can add features & performance.

AMD could also fail while adding features and performance, resulting in a chip that is larger, has lower yields, higher power consumption, and doesn't really perform better than their current line.
And what exactly did GF104 remove? They actually *added* features, it's a compute capability 2.1 card, where GF100 is only 2.0. They also made the shader dispatcher superscalar. It's not just a scaled down GF100.

 

[Caribbean Geek]

AMD's problem is the other way around:
nVidia has already figured out what to DO with the extra transistors, and has added several technologies to the GF100 architecture that AMD has yet to implement... such as a unified address space for compute shaders (giving full C++ support), ECC support, parallelized triangle setup and tessellation circuitry.

But considering that AMD always had the upper hand in the transistor density area, taking away the big serial Tessellator and breaking it into small pieces among the stream processors shouldn't be a hard task, and as far as I know, unified address space can be incorporated following a hierarchy cache model. ECC support isn't feasible for now (Look at current Fermi implementation of it), but eventually will. Current hardware have a very powerful triangle setup, by the time it becomes a bottleneck, faster hardware will be available.

 

[Scali]

But considering that AMD always had the upper hand in the transistor density area, taking away the big serial Tessellator and breaking it into small pieces among the stream processors shouldn't be a hard task

Actually, it is quite a hard task. And they don't have a *big* serial tessellator (and triangle setup engine, which is also part of it). They just have *a* serial tessellator. A 5870 is just as slow as a 5750 (and possibly even lower-end cards) at the same clockspeed in heavily tessellated scenarios. They all have the same single tessellator, there's no scaling.

and as far as I know, unified address space can be incorporated following a hierarchy cache model.

There's a bit more to it than that. nVidia's architecture also supports function pointers (in any memory). But even if AMD did get all that into their hardware, so they could support C++... they have no API or compiler to make use of it.

ECC support isn't feasible for now (Look at current Fermi implementation of it), but eventually will.

It is. ECC support is one of the reasons why companies like Dell, IBM and Supermicro now sell rack servers with GPGPU capabilities, and nVidia being their GPU of choice.

Current hardware have a very powerful triangle setup, by the time it becomes a bottleneck, faster hardware will be available.

Nope, it already is the bottleneck, see above.

 

[brybir]

/snip

As I understand it yields go down due to larger dies resulting less dies per wafer, defects per wafer, and poor design not from xtor density. Your point doesn't make a lot of sense to me.

/snip

That is because your understanding is lacking.

Making statements like this do not help people to understand whatever your point is that you are trying to make and it comes off as rude.

My answer to him would have been that IC manufacturing is a very complex process and that while hings like die size and xtor density are factors in production, there are also dozens of others issues that play a similar or more significant role in determining the actual yield results for a given chip on a given process. Even differentiating between process, synthetic and design defects that affect yield would take volumes, and people have written volumes on the topics.

 

[Scali]

Making statements like this do not help people to understand whatever your point is that you are trying to make and it comes off as rude.

I was trying to respond with a little class and tongue-in-cheek humour ("My understanding is...." -> "Then your understanding is lacking...").
Besides, I already had to repeat my previous post verbatim.

My answer to him would have been that IC manufacturing is a very complex process and that while hings like die size and xtor density are factors in production, there are also dozens of others issues that play a similar or more significant role in determining the actual yield results for a given chip on a given process. Even differentiating between process, synthetic and design defects that affect yield would take volumes, and people have written volumes on the topics.

I already said that in my previous post. Why repeat myself?

 

[Paratus]

This was all discussed in detail already.
See this thread:
Uhhhhhh... what did I just ...add? Because simple math says you are wrong.

 

[Scali]

Well according to Anand, GF104 removed most of the FP64 capability and ECC.

They may have removed ECC, we won't know for sure until they release Tesla devices based on GF104 (if they ever). ECC is only used for GPGPU, and as far as I know, only enabled on Tesla units.
As for "most of FP64"... Again, we won't know for sure until they release a Tesla device. GF100 has artificially limited FP64 performance on consumer parts as well, to make Tesla more compelling (but still FP64 performance is easily competitive with AMD's offerings).

 

[BenSkywalker]

As for increasing yield with larger dies do you have anything substantive to add? Because simple math says you are wrong.

To put it as simply as possible-

Current 460s are built so that every single one of them can have a relatively serious flaw and still be considered good. If they removed the redundant units every single chip could be bad- it is in a theoretical sense possible for them to yield 100% with a larger size die as they do it now and end up with close to 0% if they reduced the die size. It gets far more complicated then that when you start looking at why chips aren't yielding, but it is within the realm of reason to increase die size to increase yields.

Reducing the die size gives you more lottery tickets per wafer. Increasing the die size, depending on how you do it, can increase the amount of numbers you can pick per ticket.

 

[railven]

To put it as simply as possible-

Current 460s are built so that every single one of them can have a relatively serious flaw and still be considered good. If they removed the redundant units every single chip could be bad- it is in a theoretical sense possible for them to yield 100% with a larger size die as they do it now and end up with close to 0% if they reduced the die size. It gets far more complicated then that when you start looking at why chips aren't yielding, but it is within the realm of reason to increase die size to increase yields.

Reducing the die size gives you more lottery tickets per wafer. Increasing the die size, depending on how you do it, can increase the amount of numbers you can pick per ticket.

Since this too is confusing me, this post just translated to:
If you are producing a 300SP part (just a number) but are only looking for 200SP working units, than you will increase yields.

With that in my head, then they can have 100% yield results by using less "fully" functioning parts. Why bother with the fully functional parts. Isn't this the binning scheme they are using for GF100 too?

EDIT:
It's like 'this wafer is suppose to produce 400SP parts' [number again]
It has 0% yield of 400SP parts
It has 20% yield of 350 SP parts
It has 60% yield of 300 SP parts
It has 100% yield of 250 SP parts

Let's just use it for a 250 SP part and call it 100% yield!

 

[Scali]

With that in my head, then they can have 100% yield results by using less "fully" functioning parts. Why bother with the fully functional parts. Isn't this the binning scheme they are using for GF100 too?

Not really...
When you have a fully functioning part, you cannot have ANY defects in the entire chip.
When you can disable some parts, you can have some defects in the chip, and still yield a working unit.
It's all a delicate balancing act.

 

[railven]

Not really...
When you have a fully functioning part, you cannot have ANY defects in the entire chip.
When you can disable some parts, you can have some defects in the chip, and still yield a working unit.
It's all a delicate balancing act.

I know what a fully functional part is, ie my little edit.

What I'm getting at, is the chip is bigger but since they aren't harvesting "fully functional" parts from it, it can be said that the yields are higher thus going against what most would think "bigger chip, less yield per wafer."

EDIT: Sooo in summary, the bigger the chip, the more parts you can disable, the better your yield.

There you go it comes full circle! I get it now. Haha.

 

[BenSkywalker]

If you are producing a 300SP part (just a number) but are only looking for 200SP working units, than you will increase yields.

It depends. If you were going to get a 70% yield rate making the parts as 200SP units in that scenario, you would actually yield more by fabbing them as 200SP parts to start with. If you were only going to get a 30% yield rate, then the odds are you would gotten quite a bit more fabbing them as 300SP parts. It is very much a balancing act and some guess work.

It's like 'this wafer is suppose to produce 400SP parts' [number again]
It has 0% yield of 400SP parts
It has 20% yield of 350 SP parts
It has 60% yield of 300 SP parts
It has 100% yield of 250 SP parts

Let's just use it for a 250 SP part and call it 100% yield!

At the end of the day, it's all about making money. You sell the 250SP part at $100, the 300SP part at $130, and the 350SP part at $160. If your yields improve and you get something like 10% on the 400SP part you introduce that and sell it at $200. It's not just nV who does this, the reason we have the 5850 is to sell 5870 parts that don't fully yield. In this business, it is simply smart to do it this way.

 

[Scali]

EDIT: Sooo in summary, the bigger the chip, the more parts you can disable, the better your yield.

Well, sorta.
It's mainly about the distribution of the defects in the wafer.
So GF100 has 512 SPs, but only 480 SPs are used.
Now, if you were to say "Let's redesign it to be 480 SPs max".
You know you'll STILL get defects somewhere, as wafers are never perfect, and the 480 SP chip is still pretty large. So still nearly the same odds of a defect hitting a 480 SP chip.
But you now need a version with NO defects to get a 480 SP part, where you could get one with maybe 1 or 2 defects from the 512 SP design.
The 480 SP chip may turn out better for versions with more defects, like a 384 SP part. But that's hard to say.
But GF104 is probably designed more or less with this problem in mind. It's a much more economical way to get these chips out of the wafer than it is with the GF100 one...
GTX460 and GTX465 perform about the same, but GTX460 should be much cheaper to make.... It requires less defects per chip, but that is compensated partly by the smaller size per die (so less chance of a defect hitting the area), and partly by the fact that more dies fit onto a wafer.
On the other hand, GTX465 probably meant that nVidia didn't have to throw a lot away.
Things would have to go pretty bad for a chip not to able to be binned into a GTX465.

Balancing act if there ever was one

 

[Tempered81]

I bet nvidia cooks up a GF104 style gaming-centric chip with all the HPC and DP stuff removed (like GF100->GF104). Except they double or triple up on the shaders. Like a 2xGF104 part with 768 shaders and try to pump it out on 28nm Q1 '11 while keeping < 500mm^2. They could use that as a flagship gaming gpu, and would proabably best a 6870, or even a 2560sp Cayman on 28nm.

 

[Caribbean Geek]

I bet nvidia cooks up a GF104 style gaming-centric chip with all the HPC and DP stuff removed (like GF100->GF104). Except they double or triple up on the shaders. Like a 2xGF104 part with 768 shaders and try to pump it out on 28nm Q1 '11 while keeping < 500mm^2. They could use that as a flagship gaming gpu, and would proabably best a 6870, or even a 2560sp Cayman on 28nm.

Like a 1536 shader part from nVidia?, that thing would be big but definitively will be able to outperform a 25** part from AMD if its based on the same architecture as Evergreen (Like a beefed up HD 5890), but if its true that AMD stream processors are faster now on N.I., it will be a strong fight, but I'm just speculating.

 

[bryanW1995]

Well, you could say the GTX480 is under performing based on it's power use and how long it took to get to market compared to the 5870.

For using as much power as it does (more than a 5970) you would think it should out perform the 5970, but it doesn't. It's not even close.

And while there's no rule that AMD or Nvidia must bring out part within x amount of time from one another, when one does bring out new hardware for a new DX version, and the other takes 6+ months to answer, you'd expect it to be a much faster part. The GTX480 is certainly faster, but I think most were a bit underwhelmed by it. And it's not faster than AMD's fastest card. All that waiting for a part that puts out more heat, more noise, and uses more power than AMD's fastest card but is slower. See? You could certainly argue that they are under performers.

Anyway, this thread has pretty much disintegrated, it should probably be closed. I think it's run it's course. Just my $.02

actually, the last 3 times that this happened the later part sucked. 9700 pro rocked, nvidia's response was better used as a leafblower. 8800gtx was awesome, amd's response was more limp than, um, let's just say it was limp. 58x0 was good, not great, but nvidia's very late response to it was eerily similar to the others.

any time a cardmaker is this late, especially for a new directx version, slow is highly expected. imho, gf100 was bettter than anticipated because it was at least able to trumpet SOMETHING (single gpu crown), though the enormous power/heat requirements did mitigate this (partial) triumph.

 

[bryanW1995]

I have no favoritism. I buy whatever technology impresses me. I have always liked high end ATI cards, and have owned many in the past. Unfortunately I moved to a niche of clients that requires multi GPU, and for me Crossfire is simply not a competitor to SLI. I need control over my drivers, and I need to know exactly what is going on with each game, ATI does not allow me to do this. I also enjoy having CUDA and Physx, it's just something I can bank that I have in case I need it. Money is not really a concern. So you can see why I chose nvidia. Obviously if ATI had a product (let's say hypothetically the HD6870) that blew nvidia away in performance (lets say beat the GTX 480 by 50&#37; or more) I wouldn't re-consider all the above and I would purchase ATI.

What I meant to say (if it sounded contradictory) is that nvidia wasn't able to deliver a 512SP 800 Mhz air cooled GeForce card to the market. And the second problem would have been the power charts on review sites. The chip is capable of running stable with those specs, but it cannot be done for other reasons. Nothing contradictory.







You seem blind to a simple concept. This game works in rounds. One round goes to ATI, one round goes to NV. If NV has nothing until late 2011, most enthusiasts who purchased an HD6000 in early 2011, will be easily willing to sell and buy the new NV card if it delivers more performance. Remember we are enthusiasts, we want the latest and greatest. Average Joe that buys a GTX 460 and keeps it for 4 years is not an enthusiast, nor a significant source of profit in this segment.


I don't understand how you figure that NV is going to go bankrupt because the enthusiast gaming market (the smaller of their profit margin) is not buying their high end for a brief 2 to 3 quarter time span. It's like you think that "after HD6000 that's it, nobody will buy NV again, they are doomed". As long as they launch a new product that is better than the competitor's and price it accordingly, they will recap that small niche market of high-end GPU users.





Lol, thank you Seero, thank you for being the only one that puts any effort into their thinking.


And I just want to add that I'm sure this thread turned out exactly like the OP hoped for.
I'm not going to say anything else otherwise Dr Pizza will back hand me in the face.

good point about the "rounds". amd is doing well right now more because they've been able to execute on time or nearly so, while nvidia has been late and had thermal issues compared to amd. Judging by historical patterns and especially by nvidia's vastly superior financial situation it seems unlikely that they won't be able to catch up to and eventually surpass amd again, possibly in the near future.

don't forget virge, from what I hear he is very large AND he's a mod as well!

 

[Paratus]

I see where the confusion is now. I wasn't talking about disabling extra sp's to increase yields what I meant was AMD seams to be able fit 110-120 xtors in the same space that NV fits 100. From what Scali said there was some positive reason why NV chose to only fit only 100 xtors in that space.

 

[bryanW1995]

NV has predominantly used more power, you list exceptions not the rule. But hey, Im not about to go hunting the info down to show you, Im not that anal. And while you may claim people use cards not GPU's, their is still a larger faction that wont use dual GPU cards. I'm not interested in dual cards. If it was a dual core GPU, that would be different.
I still dont understand the fuss about power, in most cases its not an issue...its just something for the red team to draw attention too.
ATi make great hardware, i have never said otherwise, however their drivers suck. NV are risk takers and they have paid the price several times, however they are multi function devices, have a much better driver\software team, are more about customers rather than manufacturers and more times than not leaders in their trade.

g92 was GREAT as a shrink/low power card. in fact, it was so strong in price/performance/power usage that a lot of people actually sold their 8800gtx to get a 8800gt and cash. the 2 sides didn't really diverge in their goals until summer 2008 with 4xxx vs gtx 2xx. I would hardly list 2 generations as "the rule", especially when you consider that nvidia's latest and greatest is gtx 460 (lower power, strong performance/watt) while nvidia's upcoming gpus will be hotter/louder/larger/more power hungry than anything since...yep, you guessed it, 2900xt.

Have they released a driver that would brick a card?

try using an nvidia card for a while before bragging about amd's great drivers. until amd fixes ccc nvidia has a very strong leg up on amd in the driver dept imho.

 

[Paratus]

I've been using ATI since 1999 (RAGE Fury Pro) and never had an issue with the drivers.

Most driver stories are anecdotal.

 

[Madcatatlas]

Most driver stories are anecdotal.

or even worse, traumatic..., yes graphics cards can do that to you