AMD GPU Shortages to Continue; AMD to Add TSMC for 7nm APU/GPU 1H2018 - 2019 Due to GF Shortage

[atakall]

Original Source: Commercial Times; Repeat Source pasted below along with translation: China Times
I apologize if this is the incorrect forum. Mods, move if appropriate.
Article published 8/8/17. Human translation. Certain paragraphs are not translated due to insignificance.

http://www.chinatimes.com/newspapers/20170808000038-260202

TSMC will produce 7 nanometer APU and GPU for AMD 1H2018, 2018-2019, because Global Foundries has limited manufacturing capacity of 14 nanometer wafer.

It is expected that AMD's APU and GPU will continue to lack inventory in 2H2017 due to Global Foundries' limited capacities.

Source says that AMD will produce multiple Zen 2 APU and Navi GPU wafers next year which will use TSMC 7 nanometer. If it is smooth, it is possible that 2019 new generation 7+ nanometer APU and GPU will be OEM'ed by TSMC.

AMD revised the contract with Global Foundries for the 6th time last year 3Q, which allows AMD to purchase wafers from other manufacturers. AMD paid 100M for the deal.



Original:

據業界人士指出，超微明年將推出的數款Zen 2架構APU處理器，以及新一代Navi架構GPU晶片等，均將採用台積電7奈米製程量產。若製程轉換順利，不排除2019年的新一代7+奈米製程APU及GPU也將由台積電拿下代工訂單。



處理器大廠美商超微（AMD）今年絕地大反攻，在x86加速處理器（APU）及伺服器處理器的市占率穩健攀升，繪圖處理器（GPU）也賣到缺貨。超微為了鞏固明、後兩年產能，並且加快製程微縮速度，已決定重回晶圓代工龍頭台積電（2330）懷抱，7奈米APU及GPU訂單將在明年上半年在台積電量產投片。

超微在2014年之前的28奈米APU及GPU晶片，多數是交由台積電代工生產，2015年決定將APU及GPU產品線轉進14奈米鰭式場效電晶體（FinFET）製程後，產品線已全數交由格羅方德（GlobalFoundries）代工。因此，台積電去年及今年雖然積極爭取16奈米訂單，但來自超微的主力產品線訂單幾乎掛零。

受限格羅方德產能

超微去年推出新一代Polaris架構GPU，市場反應卻是叫好卻不叫座，主要原因就是推出之際雖然聲勢浩大，但受限於晶圓代工14奈米產能有限，無法放大出貨量，等到競爭對手輝達（NVIDIA）採用台積電16奈米製程的Pascal架構GPU推出後，超微的GPU晶圓代工產能才陸續開出。由於輝達Pascal架構GPU是較新的架構，超微等於因產能受限而失去市場先機。

下半年恐又見缺貨

超微今年以來推出一系列基於Zen架構的Ryzen處理器及EPYC伺服器處理器，目前看來銷售動能強勁，新推出的Vega架構GPU也獲得市場好評，但超微現在面臨的最大問題，仍然是格羅方德能否全力支援所需的14奈米及14+奈米產能。業界人士認為，格羅方德近年來並無擴充14奈米產能，在產能增幅有限的情況下，不排除超微APU及GPU下半年仍會出現缺貨情況。

台積續吃後年訂單？

為了解決產能限制問題，同時加快製程微縮速度，超微明年7奈米APU及GPU代工訂單將重回台積電。據業界人士指出，超微明年將推出的數款Zen 2架構APU處理器，以及新一代Navi架構GPU晶片等，均將採用台積電7奈米製程量產。若製程轉換順利，不排除2019年的新一代7+奈米製程APU及GPU也將由台積電拿下代工訂單。

超微不評論晶圓代工策略，不過，執行長蘇姿丰（Lisa Su）曾多次表示，台積電及格羅方德都是超微重要的晶圓代工合作夥伴。

超微去年第三季第6度修改與格羅方德的晶圓供應協議，其中修改條款的一條是充許超微向格羅方德以外的晶圓代工廠購買晶圓，為了爭取能下單給其它晶圓代工廠，超微支付1億美元現金予格羅方德作為代價。也就是說，超微可以尋求其他晶圓代工廠產能支援，而台積電自然是超微的合作夥伴首選。

 

[amd6502]

Is there any end to the GPU shortages in the next few months? It'd be nice to see the shortages end in the mainstream segment (RX 560- RX 480). How soon, or ever at all?

 

[ksec]

Is there any end to the GPU shortages in the next few months? It'd be nice to see the shortages end in the mainstream segment (RX 560- RX 480). How soon, or ever at all?

Note likely in the foreseeable future, unless the demand drops of coz. Supply is tight and Mobile SoC takes priorities. For TSMC that is Apple, Qualcomm and Samsung are using their Foundry. The funny thing is how GPU Polaris / Vega switched to GloFo, and now it is competing against Ryzen for capacity.

 

[JackTheBear]

It depends on what happens to the price of Etherium and the mining difficulty. Right now, miners will buy any decent AMD GPU they can get - it's going to pay for itself pretty quickly. Over time, mining difficulty goes up. If the price doesn't go up with it, mining will become less profitable or unprofitable, then there will be an enormous amount of used AMD GPUs on ebay available for cheap and you'll once again be able to buy new ones at retail stores.

 

[USER8000]

I think you are missing a crucial thing people - AMD is moving back to TSMC for its GPUs,and will moving onto a new node next year at some point. Also,we also finally can confirm like Intel they are using mobile with their latest nodes. It will be 7NM Zen2 APUs against 10NM Intel Icelake in laptops.

 

[fleshconsumed]

7nm Vega equivalent in 1H2018 sure would be nice.

 

[Yakk]

Couldn't Samsung produce if GF can't?

 

[Phynaz]

7nm Vega equivalent in 1H2018 sure would be nice.

Not gonna happen. The entire article is BS.

 

[iBoMbY]

Not gonna happen. The entire article is BS.

Yes, or at least the translation. Doesn't make sense.

 

[Ajay]

The only thing that makes sense it that AMD will move to TSMC for GPUs. Fab8 doesn't have enough capacity to produce both CPUs and GPUs, it appears.

 

[tential]

Couldn't Samsung produce if GF can't?

You produce something at the opportunity cost of something else. So why produce a low margin AMD chip, when you can produce a higher margin ANYTHING else for them?

Also just on a related/unrelated note, AMD doesn't capture the profits from higher prices at etailers. Those etailers do. I still am mindboggled when people say "It's good the cards are selling at 700+ AMD gets more money."
1. They don't, the retailer does.
2. How does that help you again?

 

[jpiniero]

^^

Then AMD should raise the MSRP so then they do.

 

[Ajay]

^^

Then AMD should raise the MSRP so then they do.

It's a no win situation for AMD, marketing-wise.

 

[HurleyBird]

Yes, or at least the translation. Doesn't make sense.

I mean, obviously you need to take the report with a grain of salt, and even if it is credible and accurate right now, things can and often do pop up in the future that push back release windows. That said, just because Vega's launch was delayed doesn't necessarily mean that other products further downrange are similarly delayed.

 

[amd6502]

The only thing that makes sense it that AMD will move to TSMC for GPUs. Fab8 doesn't have enough capacity to produce both CPUs and GPUs, it appears.

So it seems the shortage is widespread and may remain that way.

Why don't they port some GPUs and APUs to 22nm FDSOI as has been suggested by seronx? Wouldn't it be no big dea to move some 28nm tech all the way up to GCN 1.2 to such a node (assuming availability is better than fin fet)?

 

[VirtualLarry]

Why don't they port some GPUs and APUs to 22nm FDSOI as has been suggested by seronx?

Because, 14nm designs wouldn't be competitive in power usage or performance, if they were back-ported to 22nm FDSOI.

PS. Seronx should be on meds. You get used to his ramblings after a while.

 

[amd6502]

Because, 14nm designs wouldn't be competitive in power usage or performance, if they were back-ported to 22nm FDSOI.

I didn't suggest that at all. GCN 1.2 and prior is all on 28nm. >= GCN 1.3 (eg Polaris) is on FFET.

GCN 1.2 is modern and fairly current; so there's little reason to port from 14nm.

 

[darkswordsman17]

Couldn't Samsung produce if GF can't?

I believe Samsung and GFs are diverging on process tech. Samsung's 7nm and GF's are not the same. GF's is a high performance focused design, I believe targeting/aimed at high performance large CPUs. Samsung's is more aimed at lower powered efficient chips.

The only thing that makes sense it that AMD will move to TSMC for GPUs. Fab8 doesn't have enough capacity to produce both CPUs and GPUs, it appears.

That would be my guess, and I think its not just capacity, I would guess that GF's process nodes are not a good fit for GPUs. Not that that is the only issue with RTG's GPUs, but I'm sure it doesn't help and sounds like GF's 7nm is tailored even more to CPUs (although maybe it actually would be better suited for large high performance GPU since those are the biggest chips these days, and now they're working to push clock speeds up a lot). I was surprised Vega wasn't TSMC since RTG had said they were working with multiple fabs and seemed they were hinting that Polaris would be GF and Vega would be TSMC, but that was quite a while back and I don't think they said anything specific, just said they were working with multiple fabs still and that they wouldn't necessarily be locked to GF or something along those lines.

On APUs the GPU is already the bigger and more crucial aspect (efficiency, clocking, etc) there for that market so it would make sense to use the process that suits it best. Not like you'd want to be clocking your CPU as high as possible there either and you'd prefer efficiency over max performance more.

So it seems the shortage is widespread and may remain that way.

Why don't they port some GPUs and APUs to 22nm FDSOI as has been suggested by seronx? Wouldn't it be no big dea to move some 28nm tech all the way up to GCN 1.2 to such a node (assuming availability is better than fin fet)?

How would that help shortage of 14nm?

Its best to lay off his rampant bizarre speculation. You can argue about "what if AMD did ___" if you want, but they're not going to spend money porting their old architecture to a somewhat newer node. With GPUs they've moved on (and their efforts now would be better spent tweaking Vega than porting their old architecture), and APU is set to have an all new design shortly so there just wouldn't be any real benefit. AMD is doing better, but they still have to operate on limited resources and so this would just be pointless cost that doesn't play into their long term plans.

 

[VirtualLarry]

I didn't suggest that at all. GCN 1.2 and prior is all on 28nm. >= GCN 1.3 (eg Polaris) is on FFET.

GCN 1.2 is modern and fairly current; so there's little reason to port from 14nm.

If you're not suggesting reverse-porting their 14nm designs, then you are suggesting to go back and port 28nm designs to 22FDXSOI? Those are older cards, it would be mostly pointless..

My suggestion? Don't listen to SeronX. :/

 

[ao_ika_red]

Take it with a good amount of salt. TSMC's 7nm roadmap looks fishy tough. I believe their first gen 7nm is dedicated to smartphone chips. That's why Mediatek, Qualcomm, and Apple eagerly switch to TSMC. And based on Anandtech article , TSMC still has 12FFC that will be used by Nvidia in 2018. Nonetheless, if the rumour is true, (hopefully) AMD's GPU shortage will be mitigated.