No 20 nm GPUs from AMD This Year

[midnightoil]

Given that TSMC's 20nm process has been hugely delayed and is very expensive, without huge gains in density from 28nm (albeit with power draw reductions), this is hardly a surprise.

It's much more economic for both AMD and NVIDIA to wait. This has been known for ages. All the bullshit about 20nm cards (mainly NVIDIA) this year had been completely fabricated, either just idle speculation or made up to promote page views.

Add to the above that NVIDIA simply can't use the low power process (they have to use the high power process) and there's enormous competition for wafers on the 20nm low power process from ARM SoCs, and it makes absolutely no sense for either to use it this year (or perhaps at all).

AMD's next cards will either be TSMC 28nm or GF 28nm (the latter is looking increasingly likely), with the next generation definitely moving to GF on either FD-SOI and / or a smaller process.

NVIDIA's next cards will be 28nm, ones after may well skip 20nm entirely if 14nm isn't hugely delayed.

 

[cadaveca]

I think there is good reason why AMD and NVidia high-end GPUs are so similar in performance...they are using the same process. Technically, they make the exact same thing, too, with the real difference being in organization and capabilities of each part of the GPU pipeline. With AMD potentially leaving TSMC as a customer, there stands to be a real difference between AMD and NVidia offerings, and there's more to it than simply meets the eye at first glance. I am pretty eager to see that all play out and then see perhaps a reason for nVidia to start using Glo-Fo as well.

 

[midnightoil]

I think there's practically no chance of NVIDIA moving to GF. AMD may have divested themselves of their remaining stake in GF, but the two companies are still extremely close. I don't know what UMC's plans are, but they're a possible longer term option for NVIDIA, if NVIDIA can migrate their designs to a low power process (I don't think they have anything high power in the works).

But I echo your sentiment about being excited for seeing designs that aren't tuned to just one TSMC process.

 

[HumanSmoke]

I think there is good reason why AMD and NVidia high-end GPUs are so similar in performance...they are using the same process. Technically, they make the exact same thing, too, with the real difference being in organization and capabilities of each part of the GPU pipeline.

Pretty much my thoughts encapsulated. AMD have gradually integrated a more all-purpose feature set that has been de rigueur for Nvidia since G80, while Nvidia have taken a leaf out of AMD/ATI's book and looked towards efficiency (while still needing a professional workload GPU, hence the bifurcated product line) and a look-before-you-leap mentality regarding new processes. Both vendors are learning from each other- hardly surprising since both have has a measure of success in their own way.

With AMD potentially leaving TSMC as a customer, there stands to be a real difference between AMD and NVidia offerings, and there's more to it than simply meets the eye at first glance. I am pretty eager to see that all play out and then see perhaps a reason for nVidia to start using Glo-Fo as well.

I wouldn't take AMD's leaving TSMC as an automatic unless GloFo can guarantee that their process is running with good yields. If the slow ramp of Llano and Bulldozer, the lateness/yield issues with 32nm, 28nm, and the cancellation of 14nm-XM have taught us anything, it's that GloFo has its own fabrication issues

I think there's practically no chance of NVIDIA moving to GF.

How do you come by that assumption. Nvidia and TSMC signed a fab agreement in Feb 1998 for Nvidia to produce all future GPUs with TSMC, but it didn't stop them turning to IBM's 130nm FSG process for the NV 38/40/41/45 when TSMC couldn't produce guarantees over their 110nm process, and Nvidia didn't want to use TSMC's 130nm Low-K (as ATI were doing). FWIW Nvidia also used UMC (for G96/92) as late as four years ago. There is a recent history of tension between Nvidia and TSMC- some of it is posturing and actually mirrors AMD's and GloFo's public squabbles, some of it is a genuine concern over process cadence.

I don't know what UMC's plans are.

FWIW UMC were offered the same licensing deal that GloFo recently accepted.

 

[midnightoil]

How do you come by that assumption. Nvidia and TSMC signed a fab agreement in Feb 1998 for Nvidia to produce all future GPUs with TSMC, but it didn't stop them turning to IBM's 130nm FSG process for the NV 38/40/41/45 when TSMC couldn't produce guarantees over their 110nm process, and Nvidia didn't want to use TSMC's 130nm Low-K (as ATI were doing). FWIW Nvidia also used UMC (for G96/92) as late as four years ago. There is a recent history of tension between Nvidia and TSMC- some of it is posturing and actually mirrors AMD's and GloFo's public squabbles, some of it is a genuine concern over process cadence.

For the reason stated. GF are seen as, and indeed are incredibly close to AMD. There's no way NVIDIA would go to them, or not until significantly more time has passed since their separation from AMD.

I'm aware they used UMC relatively recently (didn't realise they used IBM a while ago), and they may well do again if they manage to migrate existing designs or make future designs that will work on a low power process. Hence my comment. NVIDIA's 20nm designs are all for TSMC's acutely expensive high power process, so they have no option to move to anyone else for a while, though.

AMD currently PLANS to leave TSMC entirely. That isn't automatic, as you say. But it is what they plan to do.

 

[Prima.Vera]

AMD should go to Samsung. They have the best fabs out there, and are already on 14nm. Not sure if they have the capability for producing to an industrial scale GPU's, not unless they open (a) new fab(s)...

 

[Pholostan]

Nobody is surprised really. TSMC never promised that their 20nm planar process would be able to to do high performance chips. It has been intended since the beginning for low power ARM SOC and such. Never for big GPUs. There was a rumour around that TSMC was planning a high performance branch of their 20nm planar, but TSMC themselves crushed that over a year ago. Nothing but hot air from fanboys have supported big GPUs from any maker on TSMC 20nm planar. The process was never made for it, why the hell is it news now?

 

[midnightoil]

AMD should go to Samsung. They have the best fabs out there, and are already on 14nm. Not sure if they have the capability for producing to an industrial scale GPU's, not unless they open (a) new fab(s)...

That's only memory, and is initial low volume production. As far as ARM SoCs (the other thing they fab) are concerned, Samsung's first 20nm product will be the Exynos 5430 (I think) in the international QHD version of the Galaxy S5. I don't think they're currently equipped to do GPUs or larger APUs / CPUs. So Samsung aren't really ahead of GF ... and the two are standardising their bulk FINFET low power processes anyway. Plus, with GF, if they can make FD-SOI economic, AMD have that option, too ... which could be a hell of a lot cheaper than bulk FINFET on smaller nodes.

 

[midnightoil]

Nobody is surprised really. TSMC never promised that their 20nm planar process would be able to to do high performance chips. It has been intended since the beginning for low power ARM SOC and such. Never for big GPUs. There was a rumour around that TSMC was planning a high performance branch of their 20nm planar, but TSMC themselves crushed that over a year ago. Nothing but hot air from fanboys have supported big GPUs from any maker on TSMC 20nm planar. The process was never made for it, why the hell is it news now?

There's still no word that the upcoming GPU family / families from NVIDIA will be designed for a low power process ... I was under the impression TSMC do have a high power 20nm process, specifically for NVIDIA, but it's nowhere near ready.

 

[HumanSmoke]

There's still no word that the upcoming GPU family / families from NVIDIA will be designed for a low power process ... I was under the impression TSMC do have a high power 20nm process, specifically for NVIDIA, but it's nowhere near ready.

TSMC don't have any "Nvidia only" process. The sheer cost of tooling and separate production lines (fab module) would be prohibitive in the extreme- especially for the relatively low volume Nvidia would require. The only high power 20nm planar process TSMC had on its books was CLN20G

Which was cancelled two years ago.
FWIW's all the processes being touted as 20nm successor ( TSMC's 16nm, Samsung's 14nm, Intel's 14nm) all use 20nm back end of line (22nm in Intel's case) - it's just the FEOL layer that shrinks. It is why TSMC's CLN16FF is coming hard on the heels of planar 20nm

 

[eidairaman1]

As Usual TSMC has yet another delay, they just effin suck at producing chips

 

[cadaveca]

Which was cancelled two years ago.
FWIW's all the processes being touted as 20nm successor ( TSMC's 16nm, Samsung's 14nm, Intel's 14nm) all use 20nm back end of line (22nm in Intel's case) - it's just the FEOL layer that shrinks. It is why TSMC's CLN16FF is coming hard on the heels of planar 20nm

From where I sit, TSMC and everyone else isn't exactly having a great time with current process technology, and I feel like all I hear is "we'll fix it with the next node". I'd kind of like to see everything just sit still for a moment, and for current nodes to mature a bit myself. I think about AMD's listed strategy in the past, and apply that to the console business, and I see they they have a nice seat from which they can sit back and watch the party. Nvidia is just as capable, but not due to a customer base like the consoles enjoy, but more due to sheer profitability. To me, there has never been any "mining" craze, really, and most of these current AMD GPUs have been snatched up by console devs building platforms to work from too, while Nvidia's products are headed to more mobile 3D and that platform base. So AMD's in the living room (where ATi left when dropping the VIVO designs), while NVidia is in your pocket. Intel, of course, is in business and infrastructure. Sure, there's some bleeding along those markets...but they are pretty stiff. So while conjecture about who might do what might seem interesting, provided performance greatly exceeds the needs of most users across all markets besides the mobile space, which is why we saw a push for that a while ago.

So nothing new this year? Hmm, yes please? Is this news now? I kind of understand why it would be.

 

[eidairaman1]

From where I sit, TSMC and everyone else isn't exactly having a great time with current process technology, and I feel like all I hear is "we'll fix it with the next node". I'd kind of like to see everything just sit still for a moment, and for current nodes to mature a bit myself. I think about AMD's listed strategy in the past, and apply that to the console business, and I see they they have a nice seat from which they can sit back and watch the party. Nvidia is just as capable, but not due to a customer base like the consoles enjoy, but more due to sheer profitability. To me, there has never been any "mining" craze, really, and most of these current AMD GPUs have been snatched up by console devs building platforms to work from too, while Nvidia's products are headed to more mobile 3D and that platform base. So AMD's in the living room (where ATi left when dropping the VIVO designs), while NVidia is in your pocket. Intel, of course, is in business and infrastructure. Sure, there's some bleeding along those markets...but they are pretty stiff. So while conjecture about who might do what might seem interesting, provided performance greatly exceeds the needs of most users across all markets besides the mobile space, which is why we saw a push for that a while ago.

So nothing new this year? Hmm, yes please? Is this news now? I kind of understand why it would be.

I honestly think both companies dont sit still long enough to get the best they can out of a process node, i think Intel is guilty of this aswell

 

[HumanSmoke]

As Usual TSMC has yet another delay, they just effin suck at producing chips

Well, not everyone can be Intel. Even the industry leader in process design and transistor density comes with some baggage. As for the others, Samsung ( Limitless Corruption Dept.. or LCD for short), TSMC, Globalfoundries ( King of missed new process schedules and crappy yields), and UMC (tied to Big Blue's apron strings and looking longingly over the fence at Samsung) have all either slipped schedules or are late to the party at various stages of their process evolution.

From where I sit, TSMC and everyone else isn't exactly having a great time with current process technology, and I feel like all I hear is "we'll fix it with the next node". I'd kind of like to see everything just sit still for a moment, and for current nodes to mature a bit myself.

Unfortunately, the semicon companies are in a sprint for the brass ring that is efficiency. Stopping really isn't an option when you have huge OEMs like Apple dangling huge contracts as an added incentive. While the majority of enthusiasts might just wish for fab companies to just sit pat on a given process for a while- pay off the ROI on tooling and raise yields (both lowering silicon prices), and optimize the IC's being produced to the pinnacle of what they can achieve, there is a far greater number voicing smaller node = better. How many on these forums have basically posted their dissatisfaction that the latest graphics boards are still made on 28nm and they wont buy anything unless its on a smaller process. New = Better.
Common sense doesn't seem a natural partner for the personal computer industry at the best of times. The more scrutiny you give it, the worse it gets.

I honestly think both companies dont sit still long enough to get the best they can out of a process node, i think Intel is guilty of this aswell

Intel's continued survival in processors seems linked to staying ahead of the competition in process tech. I'm not particularly sold on the idea that Intel can continue to win by going toe-to-toe with their competition on an equal footing (node size and transistor density).

 

[midnightoil]

TSMC don't have any "Nvidia only" process. The sheer cost of tooling and separate production lines (fab module) would be prohibitive in the extreme- especially for the relatively low volume Nvidia would require. The only high power 20nm planar process TSMC had on its books was CLN20G

Which was cancelled two years ago.
FWIW's all the processes being touted as 20nm successor ( TSMC's 16nm, Samsung's 14nm, Intel's 14nm) all use 20nm back end of line (22nm in Intel's case) - it's just the FEOL layer that shrinks. It is why TSMC's CLN16FF is coming hard on the heels of planar 20nm

I was under the impression it had been kept alive for NVIDIA. If not, I don't think there'll be any 20nm Maxwell products ... ever.

Could leave quite a considerable window where AMD's on a superior process, if they move to GF.

 

[eidairaman1]

Maybe next year. But lets see both companies sweat and be innovative with current process or larger

 

[alwayssts]

Intel's continued survival in processors seems linked to staying ahead of the competition in process tech. I'm not particularly sold on the idea that Intel can continue to win by going toe-to-toe with their competition on an equal footing (node size and transistor density).

I don't think anyone disagrees with that statement. There are numerous statements and business practices to support that even Intel feels this way. Since Conroe/Merom, a huge amount of their success can be distilled to their competition making mistakes and their hoarding of a process advantage at any cost. Even as things have evolved toward sub 40w tdps (mobile and low-power desktop/laptop-esque chips), their answer is now to throw what used to be leftovers onto the new (14nm) process as unique designs...just to be able to compete with what's here and coming shortly on lesser processes from companies and engineers more in-tune to achieve what people want per market.

I think we agree that if the CPA 14nm can actually get off the ground in a timely manner, Intel is going to find themselves in some very interesting fight-or-flight situations; some scenarios they haven't seen in a long time, others they have never seen before. It will be interesting to see how they react once the gap is closed and they are faced with competition from not only AMD, but others like Qualcomm whom have been forced to compete with not only less, but much less.

Once upon a time AMD openly fretted, and many journalists opined about them getting through 'the next few years'. On the other side of this transition (which seems much better organized than previous attempts) is the end of that grim period, and we should have a resurgence if well-planned. If not them, then there is ample room for a new power-player in the form of Qualcomm or even nvidia. All things considered, Intel better hold on to their butts.

Said it before, but will say it again. 2014 will likely be boring. The later part of 2015 but more importantly going into the beginning of 2016 is going to be a hugely important time, and a ton of exciting (if not unexpected) things are bound to happen across the tech industry during that year (2016).

 

[avatar_raq]

Probably the best bang for buck card I've owned in the last few years. So good that when I sold my XFX Black Editions, I "upgraded" to the Sapphire Toxic 2GB HD 5850's in my AMD machine.

5870 here, but it's showing its age performance wise and 13.4 was the last good driver, newer versions make windows 7 animations stutter for me and ruin game performance.