When is the next gen ATI/AMD Radeon due to hit? Northern Islands?

[Daedalus685]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.

Do you even try to make sense?

No, NI was designed for 28nm (32 originally)... Since it won;t be out for some time they are planning to release something that MIGHT be a half way step between NI and evergreen... or it might be true NI just at 40nm.... or it might not even exist and we will get a 5890 instead followed by NI from globalfo.


Plus, it doesn't suck for 'them'.... It sucks for people wanting to buy a graphics card made from TSMC... last I checked that was a lot of us . 28nm delays are going to bite us all in the ass with a VERY boring year if they are to be believed/global foundries is a no go for now.

 

[edplayer]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.

yeah, it does kind of suck

But they seem to have decent availability of their 5XXX series cards. Does Nvidia even have their new cards for sale yet?

and I would guess that whatever TSMC issues that ATi has would be similar for Nvidia.

 

[happy medium]

Do you even try to make sense?

No, NI was designed for 28nm (32 originally)... Since it won;t be out for some time they are planning to release something that MIGHT be a half way step between NI and evergreen... or it might be true NI just at 40nm.... or it might not even exist and we will get a 5890 instead followed by NI from globalfo.


Plus, it doesn't suck for 'them'.... It sucks for people wanting to buy a graphics card made from TSMC... last I checked that was a lot of us . 28nm delays are going to bite us all in the ass with a VERY boring year if they are to be believed/global foundries is a no go for now.

So this means the delayed 5870 refresh will be about when the gtx480 512 sp refeshes? Q4?
This could be bad news for ATI.

 

[wlee15]

I agree with this.

ATI is supposed to redo their shaders (as most applications can't use all the 4 "less capable shaders" tacked with the fully programmable shader) and whatever else for next generation so it isn't going to be a "double all shaders RV780->RV870".

That isn't true at all. The four simpler units(WXYZ) of ATI's gpu are very similar in capacity to nVidia's shader cores while the large fifth unit(T) of ATI's gpu are roughly equivalent to nVidia's SFU. The only difference is the fact the T unit can execute almost all of the same operation that the simpler units can perform.

 

[Nemesis 1]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.

Again going by your record. For instance You said Fermi be out in Sept . I have the link. So Your saying 2011 just affirms it will be out in 3rd Qt.

 

[GaiaHunter]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.

But but it's Charlie!

He is a payed fanboy that never gets anything right and when he does it is because of luck, after all a "broken clock is right twice per day!"

 

[GaiaHunter]

That isn't true at all. The four simpler units(WXYZ) of ATI's gpu are very similar in capacity to nVidia's shader cores while the large fifth unit(T) of ATI's gpu are roughly equivalent to nVidia's SFU. The only difference is the fact the T unit can execute almost all of the same operation that the simpler units can perform.

My understanding is that ATI shader setup requires the code to be written in a way that takes advantage of the 5 units and many times it doesn't.

 

[Daedalus685]

So this means the delayed 5870 refresh will be about when the gtx480 512 sp refeshes? Q4?
This could be bad news for ATI.

No, it doesn't mean that at all.

All it is saying is that NI effectively requires 28nm for them to bother with it. They can either wait for 28nm, or release something 'tweaked' this year.

Judging by the article Charlie thinks they will have a half way step SI this fall, with true NI when 28nm is out.

I don't personally think this makes a lot of sense as you cant just throw the old 'shaders' into a chip and call it a day.. If I am to believe him at all I assume he means a heavily tweaked evergreen (perhaps improving tessellation) that improves on the 5870 a good bit in the fall and a total redo in the new year for NI.

This in no way discounts a 5890 type card in the summer... in fact there are mumblings that this 'thing' (S|A forums call it a 6750, source on that was chiphell I think) has already taped out.

What we could be looking at is a drastic overhaul of evergreen for the refresh instead of a clock bump. If it actually has taped out I think the 2gb overclocked boards are all we will get from the 5870s as there would be little point to give us a 5890 in May and a new card again in July.. but who knows.


The facts as they stand are that 28nm will be late... Either ATI/Nvidia wait or they give us something. Charlie thinks the something will be a hybrid part between two generations coming out a couple quarters before the new true series is released. The new series seems to absolutely require a smaller node.. likely to preserve cost/size/etc. In any event we are all fucked.. unless aTI can pull a miracle out a new 40nm card is not going to give us much over a 5870 (I'd guess 20% at best if the size is about the same, maybe 40% if there is a lot of redundancy they can remove and the new arch is great). It will be boring for us consumers unless TSMC pulls their head out of the ass they have it in. Just think about it.. if TSMC really cant get 28nm until Q2011 we could be looking at the prices we have no for another 14months!!

 

[Dark4ng3l]

Well I think the whole late process things screws over nvidia much more than ATI as they really don't have any maneuvering room with their already too big Thermi chip. ATI can add 5-10% die space and tweak things and the chip is still going to be possible to build.

 

[Nemesis 1]

From what I have read . This is As I understand it. This is also the most practical and economic approach.
The First 6000 series will be NI but its Not the Card AMD wanted. So they decided to do the smart approach and do NI at 40 But Half of its performance.

AS I read it . 6000series NI 2 800 shader moduals with attached tess. and raster,.
3x-4x tess. improvement. 2x traiangles
225 watts 15 to 25% faster than unbroken fermi that means the 512.
This is the NI test chip for arch.

The next chip is the 28NM from Global Foundries. This is a 1 800 moduel chip to get the shrink optimized than comes the Real NI the 4 800 module 512 bus . Dubbling the perfomance of the 40nm 256 bus NI. You don't get it If NI beats Fermi2 at 40nm . 28nm well belong to AMD. The 40 nm NI IF faster than fermi is a feather in AMDs hat. Now SI could exist but these well not be 1600 shaders but 2x 800 modules there may be other features AMD brings @ 28nm

 

[yh125d]

From what I have read . This is As I understand it. This is also the most practical and economic approach.
The First 6000 series will be NI but its Not the Card AMD wanted. So they decided to do the smart approach and do NI at 40 But Half of its performance.

AS I read it . 6000series NI 2 800 shader moduals with attached tess. and raster,.
3x-4x tess. improvement. 2x traiangles
225 watts 15 to 25% faster than unbroken fermi that means the 512.
This is the NI test chip for arch.

The next chip is the 28NM from Global Foundries. This is a 1 800 moduel chip to get the shrink optimized than comes the Real NI the 4 800 module 512 bus . Dubbling the perfomance of the 40nm 256 bus NI. You don't get it If NI beats Fermi2 at 40nm . 28nm well belong to AMD. The 40 nm NI IF faster than fermi is a feather in AMDs hat. Now SI could exist but these well not be 1600 shaders but 2x 800 modules there may be other features AMD brings @ 28nm

So instead of just smaller GPUs intended to be paired up they're going with wide-scale modularity? Clusters of 800SPs.... sounds good if they can get the inter-module data transfer properly

 

[badb0y]

So this means the delayed 5870 refresh will be about when the gtx480 512 sp refeshes? Q4?

Source?

 

[Sylvanas]

My understanding is that ATI shader setup requires the code to be written in a way that takes advantage of the 5 units and many times it doesn't.

This is true. Anand has a good article on it concerning ATI's approach of Instruction Level Parallelism (ILP) and Nvidia's Thread Level Parallelism (TLP). There is a great explanation on it here http://www.anandtech.com/show/2556/6 ATI's approach has far higher theoretical performance but by nature it is not able to extract maximum efficiency all of the time. Nvidia's approach is different whereby they can keep their SM's fed with data all of the time making optimal use of their resources. So... on a SP unit basis ATI can execute an instruction in 1 clock cycle what takes Nvidia 5, but keeping those 5 units fed is difficult (as in the next clock cycle 4 may go unused which is a 80% loss in efficiency).

 

[ShadowOfMyself]

So Northern Islands is already toast and they are going with some sort of half-new architecture. This could delay things for them well into next year. Wow that sucks for them.

You are confused, this thread is not about Fermi, and we arent in 2009

 

[T2k]

Almost 100% sure it's going to be 40nm unless some magic happens outside TSMC - ATI had learn its lessons about introducing new architectures on new processes (see R520 on 90nm and chasing bug for almost a year.)

Charlie on Southern Islands.

I loooove being correct - it feels gooood.... :awe)

 

[HurleyBird]

You guys are feeding the troll.

 

[1h4x4s3x]

I loooove being correct - it feels gooood.... :awe)

SI != NI :sneaky:

 

[solofly]

http://www.nordichardware.com/en/ne...ate-in-fall-new-architecture-next-spring.html

 

[scooterlibby]

I loooove being correct - it feels gooood.... :awe)

I think it is being correct coupled with an already inflated sense of self satisfaction that feels so good, econometrically expressed as:
T2KFeelGood = B1 + B2SelfSatisfied * OccasionallyCorrect + et

 

[evolucion8]

I don't personally think this makes a lot of sense as you cant just throw the old 'shaders' into a chip and call it a day.. If I am to believe him at all I assume he means a heavily tweaked evergreen (perhaps improving tessellation) that improves on the 5870 a good bit in the fall and a total redo in the new year for NI.

At architecture level, the HD 5870 stream processors and shader layout isn't much different from the HD 4800 series or below, besides of some tweaks like more cache, GPGPU tweaks, etc., seen a diagram and you won't tell the difference except of the larger layout of the HD 5800 series. So a new SKU with Evergreen stream processors and some other stuff from NI is a nice way to test stuff like manufacturing process, technology etc.

 

[Dekasa]

Looks to me like they're almost done with Northern Islands, but will have nothing to do while they wait for the 28nm process, so they're siphoning engineers to do a hybrid stop-gap (SI) for 40nm.

No matter how it is, an extra refresh/generation can't be bad.

 

[ronnn]

looks like their good run may be coming to an end.

 

[Rezist]

From what I have read . This is As I understand it. This is also the most practical and economic approach.
The First 6000 series will be NI but its Not the Card AMD wanted. So they decided to do the smart approach and do NI at 40 But Half of its performance.

AS I read it . 6000series NI 2 800 shader moduals with attached tess. and raster,.
3x-4x tess. improvement. 2x traiangles
225 watts 15 to 25% faster than unbroken fermi that means the 512.
This is the NI test chip for arch.

The next chip is the 28NM from Global Foundries. This is a 1 800 moduel chip to get the shrink optimized than comes the Real NI the 4 800 module 512 bus . Dubbling the perfomance of the 40nm 256 bus NI. You don't get it If NI beats Fermi2 at 40nm . 28nm well belong to AMD. The 40 nm NI IF faster than fermi is a feather in AMDs hat. Now SI could exist but these well not be 1600 shaders but 2x 800 modules there may be other features AMD brings @ 28nm

I have no idea where your sources are from and I'm not saying your wrong. But a 512 bit bus will be a bigger chip then what ATi has been going for recently with there small die strategy. I don't see them abandoning a strategy that has won them the price/performance crown twice and let them dictate the market prices.

 

[Madcatatlas]

I think its only natural that things do get bigger (in some ways). Just look at the earliest gpus we had, the current ones dwarf those by miles (maaaybe with the exception of a few voodoo ones)

maybe 2xcypress on one die is "the sweet spot" two years from now.

 

[Martimus]

Another (longer) article on Southern Islands: http://www.xbitlabs.com/news/video/..._Different_Suppliers_Hybrid_Architecture.html

I'm guessing that when there's smoke, there is fire in this case, as I have seen three serperate articles on the issue, and they aren't just parroting each other.