AMD: R8xx Speculation

[dkanter]

that article also says that TSMC only has 1% of the 40/45nm capacity of F1M2. they are ramping up 32nm for production at the end of the year.

This xbit article is a bit more elaborate: http://www.xbitlabs.com/articles/other/display/globalfoundries-interview-2009_4.html

And for your lack of 40nm mentioning:


So 32/28 production should be on it's way in the end of the year. I guess that leaves plenty of space in the fabs to do 40 or.. even 45 at Dresden for R8X0.

Could you clarify what you mean by 'ramping up 32nm for production at the end of the year'?

Do you mean they will be doing wafer starts of 32nm chips? Or do you mean that chips which are produced at global foundries 32nm process will be on the market, in consumer hands?


AMD won't have any 32nm CPUs out till 2011, which is in line with typical time lines (i.e. AMD and IBM are 10+ months behind Intel to a new process node).

Based on that, I wouldn't expect any 32nm GPUs till about the same time frame.

Also, the R8xx series was way too far along to shift to GF's process. It'll be on TSMC. Remember that ATI's analog and physical design team (which does a lot of really good work - GDDR5, etc.) will need to learn the GF bulk process...which takes time and entails design risk. You definitely don't shift a project at the last minute to a new fab...

DK

 

[keritto]

It also shows 32/45nm SOI aimed at GPUs. Scattergun slide?

GPU on SOI probably wont happen except if they (ATi) wont came out with their own 64-core CPU wannabe GPU like Intel does these days

Still no graphics chip even use High-k which is more relevant even from AMDs blab marketing in time on Fusion presentation slides. So we'll probably first see that high-k metal gates, and untill then bulk is what most of GPU chips are satisfied with nowadays. Well maybe if ATi insists on that weird core concept which is surrounded by large cache and TMUs ROPs on one side. Which is even proofed flaky in current RV770 under heavy load like Furmark and probably why they prototyped quick patch on RV790 with enormous decopling caps surrounding the whole chip.
Pushing that design even future tech will have a little help on that, and not some entry level SOI. That's good for flash and z-ram ...but it seems like even amd abandoned that in favor of outsider t-ram concept.

Do you mean they will be doing wafer starts of 32nm chips? Or do you mean that chips which are produced at global foundries 32nm process will be on the market, in consumer hands?


AMD won't have any 32nm CPUs out till 2011, which is in line with typical time lines (i.e. AMD and IBM are 10+ months behind Intel to a new process node).

Based on that, I wouldn't expect any 32nm GPUs till about the same time frame.

Also, the R8xx series was way too far along to shift to GF's process. It'll be on TSMC. Remember that ATI's analog and physical design team (which does a lot of really good work - GDDR5, etc.) will need to learn the GF bulk process...which takes time and entails design risk. You definitely don't shift a project at the last minute to a new fab...

For GPUs all they need is bulk process they used all the way to K7-Barton and then initially Intel also skipped onto 130nm node before theirs 2 year tick (in some 20 month time) and AMD was just 6 month behind them maybe even they were develop it before their 24 month schedule. Anyway, I digress.
The leap behind of intel is beacuse AMDs was initially persuaded by contract with IBM, which needed SOI for theirs Power hungy monsters, to develop SOI process for IBMs needs. And they get some Flash memory advances based on it.

But before implementing SOI they have bulk process u and running at least 6 month ahead. So it's viable for them to present some 32nm low-end products. But they probably wont push too strong, cause after all with slower introduction of improved process they can renew themselves with RV880 line on 32nm node if needed
And after all TSMC 40nm works very well and if RV870 really will be only ~200mm2, it could do better than just fine on tsmc 40nm And nVidia always flirted with UMC which has more luck with 40nm shrinkage than TSMC, so i don't see why should we worry about them.

 

[dkanter]

GPU on SOI probably wont happen except if they (ATi) wont came out with their own 64-core CPU wannabe GPU like Intel does these days

GPUs on SOI will never happen, period. I agree with you.

Still no graphics chip even use High-k which is more relevant even from AMDs blab marketing in time on Fusion presentation slides. So we'll probably first see that high-k metal gates, and untill then bulk is what most of GPU chips are satisfied with nowadays.

LRB will use HKMG. Everyone else gets to wait till 32nm/28nm when TSMC and GF/IBM crack the problem.

Well maybe if ATi insists on that weird core concept which is surrounded by large cache and TMUs ROPs on one side. Which is even proofed flaky in current RV770 under heavy load like Furmark and probably why they prototyped quick patch on RV790 with enormous decopling caps surrounding the whole chip.
Pushing that design even future tech will have a little help on that, and not some entry level SOI. That's good for flash and z-ram ...but it seems like even amd abandoned that in favor of outsider t-ram concept.

GF is working with Tram, not AMD.

For GPUs all they need is bulk process they used all the way to K7-Barton and then initially Intel also skipped onto 130nm node before theirs 2 year tick (in some 20 month time) and AMD was just 6 month behind them maybe even they were develop it before their 24 month schedule. Anyway, I digress.
The leap behind of intel is beacuse AMDs was initially persuaded by contract with IBM, which needed SOI for theirs Power hungy monsters, to develop SOI process for IBMs needs. And they get some Flash memory advances based on it.

AMD used SOI because that's what IBM decided, and they were doing all the R&D. I see no reason why GF could do a bulk process any faster than an SOI one. They still need to perfect their HKMG, strained silicon, etc. get volume yields. Maybe GF can catch up a bit, but I'll believe it when I see it. They claim they'll be aggressive and catch up, but AMD's always been saying that and I don't see why it's any different now. I still expect them a year behind Intel.

But before implementing SOI they have bulk process u and running at least 6 month ahead. So it's viable for them to present some 32nm low-end products.

That makes no sense. SOI uses physically different wafers and has some different process steps as a result.

Please back up this claim with proof.

DK

 

[Razor1]

hmm didn't ATi have a notebook gpu with SOI?

 

[neliz]

Could you clarify what you mean by 'ramping up 32nm for production at the end of the year'?

Do you mean they will be doing wafer starts of 32nm chips? Or do you mean that chips which are produced at global foundries 32nm process will be on the market, in consumer hands?

They'll be doing waffer starts at the end of the year. I can't make more out of that out of this statement:

X-bit labs: It is rather well known that the vast majority of chips are produced using rather mature process technologies, e.g. 65nm and older. However, the first bulk process tech Globalfoundries will launch is 32nm tech. Isn’t such a plan too ambitious?

Tom Sonderman: We don’t believe so. 32nm bulk silicon is already running in Dresden and we’ll be ready to accept customer designs later this year with an aggressive production ramp in 2010.

So in this case TSMC seems to be lagging GF by at least a quarter assuming 32nm bulk production starts in Q4.

 

[dkanter]

They'll be doing waffer starts at the end of the year. I can't make more out of that out of this statement:

They are starting 'high risk' production. That's code for "We'll make your wafer....you may get a functional chip, two if you're lucky".

So they don't have commercially viable yields and it will probably take them 6-12 months to get there.

So in this case TSMC seems to be lagging GF by at least a quarter assuming 32nm bulk production starts in Q4.

I'd be shocked if TSMC's 32/28nm process was substantially behind IBM and AMD. Frankly, TSMC has way more volume than both combined.

And nobody is doing 'volume' 32nm this year, except Intel (which will be shipping to customers, but no systems to end users).

DK

 

[dkanter]

hmm didn't ATi have a notebook gpu with SOI?

No. Why would you ever want that?

SOI requires entirely different circuit design techniques than bulk. It costs more, and the density isn't that good.

DK

 

[Razor1]

hmm for some reason I thought they did wierd

 

[Sound_Card]

hmm for some reason I thought they did wierd

Your probably thinking of strained silicon which began with the x1700.

 

[keritto]

GF is working with Tram, not AMD.

potato - po-ta-to
Don't sell me that shark talk , you're of course right. But GF is 50% ATI stake. So what are we here talking about

AMD used SOI because that's what IBM decided, and they were doing all the R&D. I see no reason why GF could do a bulk process any faster than an SOI one. They still need to perfect their HKMG, strained silicon, etc. get volume yields. Maybe GF can catch up a bit, but I'll believe it when I see it. They claim they'll be aggressive and catch up, but AMD's always been saying that and I don't see why it's any different now. I still expect them a year behind Intel.

Well (if we sueo) i'm here with you but i never said when they'll came out with working 32nm just that's viable to see bulk 32nm alive in 1Q 2010. Considering they need some Bulldozer prototypes by the end of 2010 and it's 8-core and they will only release it on 32nm.
Yep. It's another sucker story with all that SSE5 oriented on one design instead steady development like intel SSE4.1->4.2->AES->xyz and finally 256b-AVX and that all is what 128b-SSE5 proposed in their initial specs just (with presumably only one 128b SSE engine instead two 64b they have now since Barcelona) but in 3-4 intel architectures over the years instead just one big leap on amd side, but that's another story. Simply imnsho they at least could leak out some unpolished 4-core Bulldozer with some SSE5 part 5.1-5.2-5.3

But before implementing SOI they have bulk process u and running at least 6 month ahead. So it's viable for them to present some 32nm low-end products.

That makes no sense. SOI uses physically different wafers and has some different process steps as a result.

Please back up this claim with proof.

Well that makes a lot of sense and i'm not talking they use same wafers, do i? They simply need to have working bulk, as themselves claimed, before they could orient themselves on more advanced tech at the same litography node. It's not reinventing a wheel but a simple way they adopting new lito tech. (Hint: look for Fusion blab marketing slides)

 

[rjc]

Just reporting this(cant verify): Apparently in the "wave" there are 4 amd dx11 chips currently in process. Rather than the old style RV8XX style names they have now assumed the amd based naming scheme after cities.

The 4 dx11 gpu's have been named after cities in california, the first revealed by CJ, i think is cypress. The other 3 names as yet unknown, for some reason amd is keeping this info closer to the chest than normal.

I originally doubted this when first saw it as i thought 4 gpus was too many. The market looks to be improving but is way down on where it was. I guess there is some (small) possibility the 4 could also include dx11 igp type parts or something to make up the numbers.

 

[CJ]

The 4 dx11 gpu's have been named after cities in california, the first revealed by CJ, i think is cypress. The other 3 names as yet unknown, for some reason amd is keeping this info closer to the chest than normal.

Yeah, I posted it here in February, but I heard the name already at the end of 2008. Oh. Here's another name of a city in Texas: Juniper. That's RV830 supposedly.

 

[Silent_Buddha]

The reaseon nVidia became more important for TSMC:
R8X0's produced in Dresden?

Actually the part I find more interesting from that is that Radeon x850 branded parts will no longer use the same chips as x870 branded cards.

I'm assuming they are really confident that yields will remain stellar with Rv870 (similar to Rv770) that they won't need to salvage parts and thus are going ahead with an Rv840 for x850 branded cards.

So this seems to imply that ATI/AMD will have 3 distintly different chips targeting different market segments with a 4th market segment using an x2 config all in a very short span of time.

Ambitious. It'll be interesting to see if they can pull it off.

So if rumors pan out for Dx11. Nvidia will control the enthusiast sector with ATI trying to compete there with the x2. And then ATI will have performance, mainstream, and budget?

Regards,
SB

 

[Jawed]

Including IGP in that set of 4 seems to make sense.

Cities, ugh, that's really going to be a bugger to keep track of.

Jawed

 

[Sound_Card]

Actually the part I find more interesting from that is that Radeon x850 branded parts will no longer use the same chips as x870 branded cards.

I'm assuming they are really confident that yields will remain stellar with Rv870 (similar to Rv770) that they won't need to salvage parts and thus are going ahead with an Rv840 for x850 branded cards.

So this seems to imply that ATI/AMD will have 3 distintly different chips targeting different market segments with a 4th market segment using an x2 config all in a very short span of time.

Ambitious. It'll be interesting to see if they can pull it off.

So if rumors pan out for Dx11. Nvidia will control the enthusiast sector with ATI trying to compete there with the x2. And then ATI will have performance, mainstream, and budget?

Regards,
SB

This time their is no GDDR6 to separate performance to justify a price point for a x70 sku. Most certainly we can probably count on HD 5850 being GDDR5.

 

[fellix]

Well, GDDR5 is still to implement full-differential signaling, yes?

 

[MfA]

Differential signalling is great for expansion slots, but for on PCB memory with short tracelengths it's just not worth doubling the signal pins AFAICS.

 

[dkanter]

Differential signalling is great for expansion slots, but for on PCB memory with short tracelengths it's just not worth doubling the signal pins AFAICS.

Uh no. Differential signaling lets you hit high frequencies and use reasonable amounts of power.

GDDR5 memory wouldn't be very interesting if you could only send data at 1GT/s, right?

DK

 

[MfA]

NEC has done 12 Gbps on PCB with single ended signalling.

 

[DegustatoR]

I originally doubted this when first saw it as i thought 4 gpus was too many. The market looks to be improving but is way down on where it was. I guess there is some (small) possibility the 4 could also include dx11 igp type parts or something to make up the numbers.

Is it possible that R800 (RV870x2) is getting one of those new codenames too?