TPU: AMD skipping 20nm, going straight to 14nm

[BFG10K]

Hopefully this move will help AMD stay competitive in GPU space in the future.

http://www.techpowerup.com/211973/a...-to-14-nm-with-arctic-islands-gpu-family.html

 

[dangerman1337]

Hopefully AMD will still be capable to have a good launch with a full range of HBM2 based cards and will do well sales wise (be going Pascal with a super-VR/4K build late 2016/early 2017 but maybe AMD can surprise me).

 

[ShintaiDK]

Hopefully this move will help AMD stay competitive in GPU space in the future.

Uarch is the determinator for that when both are fabless and using same nodes.

 

[dangerman1337]

Uarch is the determinator for that when both are fabless and using same nodes.

Well let us hope AMD will have a vastly improved Uarch for their 14nm cards.

 

[Silverforce11]

Uarch is the determinator for that when both are fabless and using same nodes.

Depends on which fab is responsible for making the bulk of GPUs. AMD already signed a $1B deal with GF (Samsung's process) for production wafers, which is likely to be 14nm ff due to what Lisa Su has been saying in conference calls as well was what GF is saying with their 14nm ramping.

NV was rumored to split between TSMC 16nm and Samsung 14nm. So far, TSMC is delayed. If NV is making Tegra lines on Samsung and dGPU on TSMC, they could be late to the next-node party.

Ultimately if the fabs are equal in production, then definitely its uArch as NV will also be on HBM so there's no memory advantage for AMD.

Edit: Re-worded. Typo.

 

[AtenRa]

Depends on which fab is responsible for making the bulk of GPUs. AMD already signed a $1B deal with GF (Samsung's process) for 14nm production wafers.

Where did you see that ??

 

[TemjinGold]

NV was rumored to split between TSMC 16nm and Samsung 14nm. So far, TSMC is delayed. If AMD is making Tegra lines on Samsung and dGPU on TSMC, they could be late to the next-node party.

I've no crystal ball but something tells me AMD won't be making Tegra on anything...

 

[Chicken76]

...If AMD is making Tegra lines on Samsung...

AMD makes Tegra?

 

[Silverforce11]

Typo.

 

[guskline]

Good news for AMD. Smart move.

 

[brandonb]

For all I care, they can take their time (NV too).

In the past, we used to have games that you couldn't run on the latest hardware, and you needed the new video card. But those days are long past us, and now games are made for hardware of 4 years ago.

These days if they released the same product as my 7970, gave it DX 12 feature levels, and made it more power efficient, I'd buy it.

 

[frozentundra123456]

Hasn't it long been known that there were going to be no 20nm dgpus?

 

[ShintaiDK]

Hasn't it long been known that there were going to be no 20nm dgpus?

Yep. But that doesnt make people click sites today to keep the ads flow running. So a new story is made.

 

[Cloudfire777]

Depends on which fab is responsible for making the bulk of GPUs. AMD already signed a $1B deal with GF (Samsung's process) for production wafers, which is likely to be 14nm ff due to what Lisa Su has been saying in conference calls as well was what GF is saying with their 14nm ramping.

NV was rumored to split between TSMC 16nm and Samsung 14nm. So far, TSMC is delayed. If NV is making Tegra lines on Samsung and dGPU on TSMC, they could be late to the next-node party.

Ultimately if the fabs are equal in production, then definitely its uArch as NV will also be on HBM so there's no memory advantage for AMD.

Edit: Re-worded. Typo.

Another soul grasping on any hope but this 14nm for discrete GPUs which wccftech reported today is surely wrong.

Look, Lisa Su said in the conference call that they will be buying wafers worth 1 billion from GlobalFoundries.
Those are 28nm SHP for 300 series and Carrizo which is the only products she mentioned will be coming, plus APUs for the video consoles (Xbox One and PS4) of course.

She didnt specify 14nm anywhere and none of AMD products are applicable for 14nm FinFETs. That process is Low Power only.
Global Foundries have 8 fabs and only 1 of them are allocated to 14nm, the rest (3) are from Samsung. Because its for mobile and low power only which Samsung is expert of. The majority of GloFo fabs are working on 28nm (maybe 20nm too) because thats what AMD have ordered for their chips.

 

[RussianSensation]

Another soul grasping on any hope but this 14nm for discrete GPUs which wccftech reported today is surely wrong.

I don't think anyone expects the current generation of Maxwell/R9 300 series to be on 14nm. However, future generations will adopt 14nm/16nm nodes with HBM2, rumoured for 2016 (likely late 2016). This is what this news is about.

 

[Cloudfire777]

I don't think anyone expects the current generation of Maxwell/R9 300 series to be on 14nm. However, future generations will adopt 14nm/16nm nodes with HBM2, rumoured for 2016 (likely late 2016). This is what this news is about.

I don`t think there will be 14nm for GPUs. First we will have 16nm from TSMC in 2016 and atleast 2-3-4 years. Then we will get 10nm from TSMC.

Unless GloFo comes up with a process that can do High Power imo

 

[Buttercream]

TSMC is planning on 16nm finfet production in Q3 this year. Then 10nm in 2016, then 7nm in 2017.

Whether they can stick to the schedule is doubtful.

 

[RussianSensation]

I don`t think there will be 14nm for GPUs. First we will have 16nm from TSMC in 2016 and atleast 2-3-4 years. Then we will get 10nm from TSMC.

Unless GloFo comes up with a process that can do High Power imo

Well since most of us aren't engineers, what we really care about is the key theme here. Whether AMD will have GPUs on 14nm or 16nm is basically irrelevant. The key theme here is that next generation node (14nm OR 16nm) with a brand new R9 400 series using HBM2 is rumoured to be on schedule for late 2016. That means R9 390X is going to be a short-lived generation like HD6970 vs. HD7970 was. I think that's the major takeaway from this news.

HD6970 - Dec 15, 2010
HD7970 - Dec 22, 2011 (new node + new architecture = massive increase in performance, which today really approaches 75-100% in modern titles).

A lot of people on this forum are of the view that the move to HBM2 + lower node (14nm or 16nm) and a new architecture from NV (and maybe AMD) will mean one of the biggest leaps in GPU performance in 1 generation. The last time NV did this, GTX780Ti beat GTX580 by 2X. If we look at Titan X vs. 780Ti, it's about 56% at 4K. I wouldn't be surprised if GP200 will be 75-100% faster than the Titan X because NV is bringing HBM2 + new architecture + new node. It's really hard to ignore the significance of all 3 of these changes simultaneously and for that reason I expect a 550-600mm2 14-16nm NV chip to be a bigger leap than Kepler --> Maxwell. The only way I don't see a huge leap is if NV intentionally splits the generation again and first gives us GP204, then a cut-down GP200, then a fully unlocked GP200 and then a fully unlocked + max clocked GP200.

I pretty much expect a Pascal's mid-range GTX980 successor to easily beat the Titan X by Q4 2016. AMD shouldn't be far behind either.

 

[Riceninja]

sounds like the r400/pascal cards are coming in about a years time. that makes the R9 300 generation rather short lived given that it won't even be announced til june.

 

[njdevilsfan87]

sounds like the r400/pascal cards are coming in about a years time. that makes the R9 300 generation rather short lived given that it won't even be announced til june.

I hope it is given how long we've been on 28nm.

 

[NostaSeronx]

Ya, cause 14nm FinFETs and 14nm Planar FDSOI aren't actually 20nm FinFETs and 20nm Planar FDSOI. It would be more interesting if AMD was skipping 20/14nm for 10nm which is true 14nm. Weird marketing is just weird.

 

[Subyman]

The only way I don't see a huge leap is if NV intentionally splits the generation again and first gives us GP204, then a cut-down GP200, then a fully unlocked GP200 and then a fully unlocked + max clocked GP200.

IMO there isn't any doubt they will do that. Nodes lasting so long practically require them to do business that way. I would be very surprised if nvidia threw that much performance onto the market all at once. That is the kind of thing you do when you are trying to bury competition, not when simply compelling owners of 3 year old cards to upgrade. AMD is of no practical concern to Nvidia anymore. Bumping prices down $50 when AMD launches cards competing with cards Nvidia has enjoyed 6 months of no response to is all the thought they give them any more. Sad to see.

Nvidia has tons of potential response up their sleeve for whatever AMD throws out there. Without AMD, nvidia will trickle performance onto the market in scheduled releases, much like Intel.