The AMD Execution Thread [2007 - 2017]

[Squilliam]

Oh indeed, I think even if Fermi blazes the trail and outperforms 5970 it won't have much of an impact as the chip will probably be in short supply for quite a while, probably for the whole if its initial existence as a 40nm part.

Oh yes, definately. If chip size, number per wafer and number of viable per wafer are anything to go by.

The point you make about mainstream parts is important on two fronts, one as you point out they make up most of the dollars spent on discrete graphics and therefore the profit to be earned. Second is the timing, you say Q3 new architecture (as opposed to refresh), it would be very difficult for ATi to introduce a volume product on 28nm on a completely new architecture in Q3 (when TSMC say they go into phase 5 production) even on a refresh it would be quite hard and maintain reasonable yields/costs. If Q3 is their stated goal [for a refresh] it would make sense to do 40nm at massive volumes/high yields (which are apparently available now) and wait until Q1/2 2011 for a new architecture on volume 28nm. Add in that Nvidia are probably hammering TSMC for all of their 28nm availability right now (maybe even willing to pay a lot per wafer, more than ATi at least) because the attraction of moving down to 28nm must be huge for Nv given the heat/size/cost concerns Fermi has.

Their architecuture is a good one, its giving them solid performance per watt, per mm^2. However it is an old one and it looks like Cypress is about as far as they can take it on the current design parameters. its on the high end where they need their refresh. The rest of the product range is fine/nigh untouchable at this stage. If they need to get Fusion out in 2011 on SOI, then they may be looking at getting one of their architectures primed for 32nm SOI production this year. So even if they don't come out with a new architecture this year, they'll certainly come out with a refresh on GloFo's 32nm process. GloFo is the reason why TSMC cannot simply give all their 28nm resources to Nvidia and its also why ATI won't be stuck if they do.

Given the architectural requirements the best chip for a 32nm refresh would be the HD 5870(Cypress) as on 32nm they could finally make it fit for latop production under the 59xx branding. The architectural changes needed to make the chip fit for Fusion would also prolong its life because you have to consider the changes they would have to make to the cache/register structure to enable the unit to live with higher potential memory to GPU latency and lower bandwidth, which would also scale extremely well for the desktop part as well.

It wouldn't surprise me if Nvidia's mainstream part never makes an appearance at 40nm being 28/32nm only, with reportedly poor yields at 40nm for GF100, the mainstream won't be that much better, not double anyway. So moving to 28nm and riding out the storm of little availability in Q4 whilst denying ATi 28nm for all 2010 would be a something I would pursue as Nv.

I would be surprised. They need them, so they ought to be coming. G80 based parts cannot compete with Juniper forever. They would also be pretty embarrassed to cede that much market share.

Basically, it looks like GF100 is so uneconomical right now even paying over the odds for 28nm per wafer might be cheaper for Nv per usable die and give better heat/size/cost than 40nm. While Cypress is anything but uneconomical and competing with Nv for 28nm on Nv's terms probably won't suit them, not until volume outstrips supply early in 2011 and TSMC can give lower prices.

I wouldn't count those birds until they are in the hand. Theres no reason to assume that 28nm would be perfect for them when they had trouble with 40/55nm.

Though I wouldn't advocate that ATi sit back on the success of Cypress and the apparent failure of Fermi, it looks like they can rake in a lot of dough between now and Jan 2011 and move to 28nm with a brand new architecture fully tested and ready for mass production while Nv are struggling to make 28nm work on the current architecture.

I hope so!

I don't know, just thinking out aloud...

Me too.

 

[Bouncing Zabaglione Bros.]

The point you make about mainstream parts is important on two fronts, one as you point out they make up most of the dollars spent on discrete graphics and therefore the profit to be earned. Second is the timing, you say Q3 new architecture (as opposed to refresh), it would be very difficult for ATi to introduce a volume product on 28nm on a completely new architecture in Q3 (when TSMC say they go into phase 5 production) even on a refresh it would be quite hard and maintain reasonable yields/costs. If Q3 is their stated goal [for a refresh] it would make sense to do 40nm at massive volumes/high yields (which are apparently available now) and wait until Q1/2 2011 for a new architecture on volume 28nm. Add in that Nvidia are probably hammering TSMC for all of their 28nm availability right now (maybe even willing to pay a lot per wafer, more than ATi at least) because the attraction of moving down to 28nm must be huge for Nv given the heat/size/cost concerns Fermi has.

There's also the possibility that AMD might be looking to switch to Global Foundries for 28nm - something which GF is apparently well ahead of compared to TSMC. I don't see NV buying up all production capacity at two foundries just to keep AMD out of the game, especially as at GF AMD also carries the weight of their previous relationship and their CPU orders. I don't think NV will be in a good place for the next 12 months. They either have to sell an uneconomical chip, not sell it at all, or have to go ahead with a very new (and knowing TSMC, somewhat unreliable) redesign for 28nm.

It's currently AMD's race to lose, and they could be transitioning to 28nm with new products as soon as possible in order to keep a process lead over NV for another 6-12 months from Q3/4, repeating what happened in Q3 09.

 

[NathansFortune]

Their architecuture is a good one, its giving them solid performance per watt, per mm^2. However it is an old one and it looks like Cypress is about as far as they can take it on the current design parameters. its on the high end where they need their refresh. The rest of the product range is fine/nigh untouchable at this stage. If they need to get Fusion out in 2011 on SOI, then they may be looking at getting one of their architectures primed for 32nm SOI production this year. So even if they don't come out with a new architecture this year, they'll certainly come out with a refresh on GloFo's 32nm process. GloFo is the reason why TSMC cannot simply give all their 28nm resources to Nvidia and its also why ATI won't be stuck if they do.

Fusion is a bit of a wildcard tbh, I wouldn't put too much stock into it competing with discrete solutions like Cypress or Fermi. Agreed on GF, though how much influence AMD retain there is another wildcard as Nv could very well go there for 32nm as well for mainstream in Q3/4. Money is money after all. If AMD can block Nv's use of GF while they still maintain a stake in the company, all the more reason for Nv to book up as much 28nm capacity at TSMC.

Given the architectural requirements the best chip for a 32nm refresh would be the HD 5870(Cypress) as on 32nm they could finally make it fit for latop production under the 59xx branding. The architectural changes needed to make the chip fit for Fusion would also prolong its life because you have to consider the changes they would have to make to the cache/register structure to enable the unit to live with higher potential memory to GPU latency and lower bandwidth, which would also scale extremely well for the desktop part as well.

Agreed.

I would be surprised. They need them, so they ought to be coming. G80 based parts cannot compete with Juniper forever. They would also be pretty embarrassed to cede that much market share.

G92-4lyf yo...

I don't think it will matter that much, Nv have proven just how willing they are to rename, so why not one more for a GT320/30 part just to have a presence on the market until Q3 when Fermi derivatives are ready.

I wouldn't count those birds until they are in the hand. Theres no reason to assume that 28nm would be perfect for them when they had trouble with 40/55nm.

Without a doubt, I wouldn't suggest that, I was just pointing out that with the same or even slightly more awful yields Nvidia have a lot to gain from the move down to 28nm. GF100 is probably going to go down as the most uneconomical GPUs ever released, it's like Nv didn't learn their lesson from the GT200 debacle of being too big/hot/expensive. The only saving grace for GT200 was that it was early, at least earlier than RV770.

I hope so! .

You can see what I'm getting at, Nv have a lot more to gain and less to lose from chancing 28nm than ATi do. If the transition for ATi goes unexpectedly poor, it leaves an easy way back into the market of viability for Nv. For the first time in a ages ATi have the upper hand, and they need to play their cards at the right time.

Working on getting RV970 to decent yields at 28nm should be their first port of call so they can have a massive launch in 2011 with no shortages like RV870 had.

Obviously this is very dependent on Nv not pulling a last minute rabbit out of the hat and Fermi turning out to scale down to 28nm well and mainstream parts coming out easily (which for the record I think won't happen, Nv have a lot of pain to go through yet in the discrete sector to remain viable).

Anyway, neither company is about to go bankrupt, Nvidia have got a massive future lined up in the embedded sector with Tegra2, and Ion 2 should (hopefully, I want a new netbook!) start to trickle out in the next couple of months while AMD look like becoming profitable again.

 

[rpg.314]

What makes people here think that gf100 won't be shrunk to 28 nm at the first sight of 28nm trannies?

NV has learned it's lesson with the gt200 hammering. IMHO, they are not gonna let amd have the process lead anymore.

 

[aaronspink]

There's also the possibility that AMD might be looking to switch to Global Foundries for 28nm - something which GF is apparently well ahead of compared to TSMC. I don't see NV buying up all production capacity at two foundries just to keep AMD out of the game, especially as at GF AMD also carries the weight of their previous relationship and their CPU orders. I don't think NV will be in a good place for the next 12 months. They either have to sell an uneconomical chip, not sell it at all, or have to go ahead with a very new (and knowing TSMC, somewhat unreliable) redesign for 28nm.

It's currently AMD's race to lose, and they could be transitioning to 28nm with new products as soon as possible in order to keep a process lead over NV for another 6-12 months from Q3/4, repeating what happened in Q3 09.

People are WAY WAY too optimistic about 28/32 nM from the foundries. Have you people learned nothing from history? You expect GF/TSMC to hit on their 28nM process before Intel has moved to <= 28nM?

Intel has basically JUST released 32nM products to market. Generally it takes GF/TSMC 12-18 months before they hit real production on the roughly the same geometries, yet people think that not only will GF/TSMC hit production before Intel, but also on a smaller geometry. TSMC has NEVER been able to though they've talked about it for the past 4-5 generations. Same goes for AMD/GF. I just don't see what has fundamentally changed at either company to allow them to do it.

 

[itsmydamnation]

Oil money.........

 

[nAo]

Oil money.........

Is there anything that oil money cannot buy!?

 

[rpg.314]

People are WAY WAY too optimistic about 28/32 nM from the foundries. Have you people learned nothing from history? You expect GF/TSMC to hit on their 28nM process before Intel has moved to <= 28nM?

Intel has basically JUST released 32nM products to market. Generally it takes GF/TSMC 12-18 months before they hit real production on the roughly the same geometries, yet people think that not only will GF/TSMC hit production before Intel, but also on a smaller geometry. TSMC has NEVER been able to though they've talked about it for the past 4-5 generations. Same goes for AMD/GF. I just don't see what has fundamentally changed at either company to allow them to do it.

Well, for all TSMC's problems, they did ship 40 nm products (in millions) in the Q4-10 while Intel was on 45 nm.

And just because tsmc borked the 40nm node deosn't mean they can't get it right the next node. Sure, Intel has had the process lead for a long time now but that is not a law of nature. Overtaking intel will take a lot of effort, no doubt. But it can be done.

 

[silent_guy]

Well, for all TSMC's problems, they did ship 40 nm products (in millions) in the Q4-10 while Intel was on 45 nm.

The fact that Intel never cared for and thus never tried to do 40nm, makes your statement a little irrelevant, don't you think?

And just because tsmc borked the 40nm node deosn't mean they can't get it right the next node. Sure, Intel has had the process lead for a long time now but that is not a law of nature. Overtaking intel will take a lot of effort, no doubt. But it can be done.

Good material for a motivational speech but they don't change the fact that Intel can easily outspend their competitors in R&D and that they simply have the best talent in process engineering. As long as that doesn't change, "a lot of effort" is just an empty slogan. It will take a lot of effort just to match Intel's technology a year later.

 

[aaronspink]

Well, for all TSMC's problems, they did ship 40 nm products (in millions) in the Q4-10 while Intel was on 45 nm.

They started shipping their 40nm at the same time that Intel was delivering 10s of millions of 32nm parts to OEMs.

And just because tsmc borked the 40nm node deosn't mean they can't get it right the next node. Sure, Intel has had the process lead for a long time now but that is not a law of nature. Overtaking intel will take a lot of effort, no doubt. But it can be done.

possibly but TSMC has a pretty long track record of "borked" processes that generally result in them actually being able to do real volume manufacturing 10-12 months after they claim they can. 40nM is no different.

 

[aaronspink]

Oil money.........

Yep cause oil money knows so much about process technology. Oh, wait. Oil money isn't some magic fix all. Its would take LOTs of oil money and more importantly LOTs of years to catch up in cadence. And quite honestly, I don't think the oil money's plan is to just keep dumping more and more money in. They do want an actual return on investment.

 

[chavvdarrr]

a bit side note:
Is anyone aware if Radeon's higher density per mm2 is "free"?
I heard rumours that the high tran/mm2 density makes R8xx a bit more expensive to produce.
Like 80-90$/chip now with supposed price of Gf100 at 120-130$/chip

 

[aaronspink]

a bit side note:
Is anyone aware if Radeon's higher density per mm2 is "free"?
I heard rumours that the high tran/mm2 density makes R8xx a bit more expensive to produce.
Like 80-90$/chip now with supposed price of Gf100 at 120-130$/chip

As long as it follows the process rules it shouldn't really make any difference. The transistor density of the design comes from having minimal control per alu. The also has the downside of the 770/870 having generally lower execution efficiency do to all the restrictions/issue with a VLIW design.

 

[rpg.314]

They started shipping their 40nm at the same time that Intel was delivering 10s of millions of 32nm parts to OEMs.

OK

possibly but TSMC has a pretty long track record of "borked" processes that generally result in them actually being able to do real volume manufacturing 10-12 months after they claim they can. 40nM is no different.

I didn't know that. Will keep an eye on it from now on.

 

[itsmydamnation]

Yep cause oil money knows so much about process technology. Oh, wait. Oil money isn't some magic fix all. Its would take LOTs of oil money and more importantly LOTs of years to catch up in cadence. And quite honestly, I don't think the oil money's plan is to just keep dumping more and more money in. They do want an actual return on investment.

for someone so partionising your making just as many assumptions as my two words . TSMC has been "miss firering" for a while, pooring in extra money into GF right now to help beat TSMC to market as well as having a better producted would be a big win that should pay off over the life time off the process also creates a halo effect for future processes.

We also have no idea how far along intels 32nm process is/has been. it could have been ready for 6/12/18 months and waiting for designs to be completed etc. We also know that a CPU takes a lot longer to come to market, look how long people have had mangy corus ES's its still not coming out for another 2 months. look at nelham they had to go to C0 and it still took until D0 to get it cool. compare that the GPU's and the lead time intel has over the compitition can apear a lot shorter then it actually is if TSMC/GF deliver.

 

[aaronspink]

for someone so partionising your making just as many assumptions as my two words . TSMC has been "miss firering" for a while, pooring in extra money into GF right now to help beat TSMC to market as well as having a better producted would be a big win that should pay off over the life time off the process also creates a halo effect for future processes.

The problem is you don't understand the technology. While designing processes and building fabs take some substantial financial resources, you cannot just dump money in and expect a miracle. Its a marathon and not a drag race. Process design and commercialization takes a significant amount of time. The timescales are realistically on the order of a decade, not months. FYI, the current trend is that Intel is on a faster cadence than anyone else except samsung in process technology. Not only are those two ahead, they are pushing forward as a faster rate than the competition. Some of this is research/people based, some of it is market based, some of it is not being a foundry, and some of it is increased resources. The best case realistic outcome is that increased resources would keep GF on the same cadence as Intel/Samsung, it would take major missteps by either Intel or Samsung for most of the other semiconductor manufacturers to close the actual gap. And FYI, TSMC actually has more resources than GF as far as process research. They are realistically only behind Intel/Samsung in people, management, and resources.

But feel free not to take my word for it, I'm just an interested party whose been following the process race for 15+ years. GF and AMD before it has pretty much been a small fry in the world of semiconductor technology and their hookup with IBM on process research had a lot of people scratching their heads as IBM semi plays in a different quadrant than amd should be playing in when it comes to process technology.

We also have no idea how far along intels 32nm process is/has been. it could have been ready for 6/12/18 months and waiting for designs to be completed etc. We also know that a CPU takes a lot longer to come to market, look how long people have had mangy corus ES's its still not coming out for another 2 months. look at nelham they had to go to C0 and it still took until D0 to get it cool. compare that the GPU's and the lead time intel has over the compitition can apear a lot shorter then it actually is if TSMC/GF deliver.

You are comparing products with processes which is an incorrect comparison. Also, you are comparing significant architectural design changes to minor evolutions. Intel historically has had their shrinks ready for volume production inline with process volume production.

And FYI, GPUs are currently at roughly the same timelines in post-si as CPUs for equivalent levels of changes.

 

[2senile]

hookup with IBM on process research had a lot of people scratching their heads

Wouldn't it make sense for materials research. Eventually Silicon will have to be replaced & I'm betting AMD never had the cash to do that sort of research on their own.

 

[aaronspink]

Wouldn't it make sense for materials research. Eventually Silicon will have to be replaced & I'm betting AMD never had the cash to do that sort of research on their own.

nope, all the research was effectively for the next 2 generations or so. the issue is that IBM is primarily focused on their high end chips and not focused on cost. Don't get me wrong, at any given point IBM's process is very good, but it also is fairly expensive.

 

[overclocked_enthusiasm]

Interesting to see how a three year old thread I started has evolved after not posting anything at B3D in over 2 years! AMD has certainly gone through some major changes and ATI has recovered nicely in AMD's house after a rocky start. I think a new balance has been struck between Intel and AMD with the recent settlement and now AMD will be allowed to survive as second fiddle to Intel in CPUs. In GPUs I think ATI/AMD will lead the way and perhaps in the future a true Fusion will take place.

 

[Kaotik]

Good question was raised at muropaketti - is there any actual source for TSMC skipping 32nm bulk process, excluding the usual rumormills at fud/bsn/sa?