(8-17-13) AMD 9970 28nm vs. 20nm Poll

[BallaTheFeared]

I'll believe that when I see it :|

 

[AnandThenMan]

Volume 20nm doesn't begin until Feb/2014, so unless AMD plans on having a select few boards and wants you to wait 6+ months to buy the 9xxx cards....

 

[tviceman]

Assuming it releases this year, which it likely will be based on the fact they're having a press event in September, then absolutely positively without a doubt no way it will be 20nm.

 

[tviceman]

The only thing you're gonna be disappointed in is Nvidia's 20nm response in 9 months time.

Volcanic Islands was built for 28nm at it's most mature stage because AMD knows how TSMC will perform better than TSMC does, having been let down by them so often in the past.

The cards have probably been ready for a month or two already, but AMD decided to wait on BF4. This holiday season Nvidia is going to face their worst nightmare with the new AMD cards having a big lead and an amazing Never Settle bundle.

All the talk has been about how Never Settle is costing AMD money, but you guys just don't realise how long these things take to play out.

LOL ok. Every part of this comment is over-the-top laughable. AMD is not going to have a "big lead" over nvidia this year. This is exactly the same crap that was spouted with Cayman. How it was ready but being held back, was supposedly 30-40% faster than a gtx480, was the next coming of Jesus, blah blah blah....AMD will manage to tie or slightly overtake a gtx780, which Nvidia will easily be able to manage. I'm bookmarking this post so I can link it on 8970 release day.

 

[BallaTheFeared]

True true, but if you had nothing to counter the 780, what would you do?

Talk about whats next, or sit quitely in the corner while your $550-600 product slowly dwindled to sub $300 markets?

I'm not convinced AMD has a new 28nm product refresh, I think we're past the point of refreshing 28nm, it's been ages since Nvidia did their refresh and while AMD had been later in the past, not this much, and not this close to a new node which they've been well to grab first over Nvidia.

I could very well be wrong, buy my intuition says this date sets a time when AMD will publicly divulge information about a product that will obsolete the 780. But it won't be a product released this year.

That is my guess, 28nm refresh, two years later, on an ancient node after saying they were sticking with 7xxx throughout 2013, to me it makes no sense. Not with 20nm months away, and AMDs ambition to be first to market on new nodes. That would make the cycle time on a 28nm product insanely short, coupled with the need for a larger die, it makes no sense financially to me.


What does AMD gain from contesting Nvidia this late in the node on a chip that had considerable financial backing in the professional market which AMD knows they won't get?

AMD doesn't have a ton of resources to toss around, they're not the 800 pound gorilla in the room. They need to play it smart, and a 28nm refresh a few months out from 20nm ramp up isn't a smart move imo.

 

[tviceman]

If it's 28nm that might indicate that 20nm is not going to be here very soon, rather than being "stupid late".

Or that this architecture was delayed with the layoffs AMD had over a year ago, or that 20nm is still going to be here in 9-12 months time but that this was AMD/s plans/intentions all along.
I don't think that there is any reason at all to look into new 28nm GPU's being released right now as a sign of something. It just is what it is. Nvidia released GT200 on 65nm when 55nm was 100% ready to go, so it's not unprecedented for a company to wring out as many products as possible on a proven node process before moving on.

 

[tviceman]

I could very well be wrong, buy my intuition says this date sets a time when AMD will publicly divulge information about a product that will obsolete the 780. But it won't be a product released this year.

If talking about compute, then ok I can see where AMD would want to raise some eyebrows and make some noise in an attempt to gain any sort of opening in the professional space. If it's gaming, then talking about a product in September that will beat your competitor's product, let alone destroy your own existing products, when it isn't even coming out for 4+ months is a sure fire way to kill sales of your existing lineup.

Cayman came out almost exactly 12 months before Tahiti. AMD laid off GPU engineers a little over a year ago. A small delay of a month or two (because of lay offs) would have otherwise put this product right in the same time frame or slightly ahead of Cayman in respect to time out before GPU's on a new node came out.

AMD doesn't have a ton of resources to toss around, they're not the 800 pound gorilla in the room. They need to play it smart, and a 28nm refresh a few months out from 20nm ramp up isn't a smart move imo.

I think 20nm GPU's are coming right at this time next year. August / September. I think Apple and Qualcomm will buy up all the initial months of 20nm wafer supplies before Nvidia and AMD afford to jump in.

 

[BallaTheFeared]

But the question is... Is it beneficial for AMD to attack the 780 given it's high price and medial market allotment this late in a node?

I think AMD is sitting fine, they don't need to contest a $650 product there isn't a ton of money at that bracket in the first place.


Plus you have to consider timeline and cost to market, if they're not releasing in September, which we're pretty sure they aren't that leaves us very close to the end of the year. All reports I've seen show 20nm ramp up in the latter half of Q1 beginning of Q2 which puts AMDs refresh 3-6 months prior to 20nm availability.


Granted we expect Nvidia to be late as always, but what advantage does AMD gain going after the 780? They have a good product stack that was tarnished by poor driver support. They're competitive from $400 on down, what good is the $400+ market for AMD from a financial standpoint?

How do they benefit from spending money refreshing 28nm now, only to short it to hit 20nm, or be lane on that node?

 

[Zanovar]

This is what has me confused,what is the point this late in the day 28nm?and if it is 28nm which seems likely a snoozefest:S.battling with a 780?*yawn*

 

[tviceman]

But the question is... Is it beneficial for AMD to attack the 780 given it's high price and medial market allotment this late in a node?

I think this is just AMD's product cycle cadence - to release two entirely different product lines per node. They did it at 40nm, and now they're prepping to do it at 28nm. I think AMD can get 780 performance (or ever so slightly better) out of a 385-400mm^2 GPU and sell it for $550 and make just as much money per card as Nvidia does off the gtx780. But AMD has said they are going to be basing much of their future GPU tech off of GCN (improved with each iteration, of course). I don't think this will be as much of a change from Tahiti as Cayman was from Cypress. I think we'll be seeing a tweaked GCN architecture with some efficiency improvements (I'm predicting ever so slightly higher power draw than a 7970GE and a ~3-5% better performance than a gtx780 in a large suite of gaming benchmarks).


Think about these musings for a minute:
I posted awhile back that AMD's engineer layoffs last year in conjunction with their aggressive GPU product cadence was going to catch up with them eventually. We may be seeing the first signs of that now (although, like I said in my previous post, releasing new GPU's in September about a year out from when I think 20nm GPU's will be released falls right in line with what happened at 40nm). At 40nm AMD released 2 different GPU families with significant architecture overhauls. Nvidia released a cripple Fermi then fixed Fermi. At 28nm Nvidia released Kepler and only Kepler, whereas AMD is shaping up to release GCN and an overhauled GCN. Keep in mind the more transistors Nvidia and AMD pack into their GPU's, the more difficult these GPU's are to make which may also be another reason why so many people feel like these GPU's are coming out "late."

 

[Anarchist420]

If it's gaming, then talking about a product in September that will beat your competitor's product, let alone destroy your own existing products, when it isn't even coming out for 4+ months is a sure fire way to kill sales of your existing lineup.

The merchants have already bought them.

 

[ShintaiDK]

AMD doesn't have a ton of resources to toss around, they're not the 800 pound gorilla in the room. They need to play it smart, and a 28nm refresh a few months out from 20nm ramp up isn't a smart move imo.

This is what has me confused,what is the point this late in the day 28nm?and if it is 28nm which seems likely a snoozefest:S.battling with a 780?*yawn*

20nm is not even close. Specially not for GPUs. So they need a refresh to keep selling cards. Else you are essentially telling them to earn nothing for the next 9-12 months. Hell, if we are real unlucky we might first see 20nm GPUs in 15 months time.

 

[raghu78]

Plus you have to consider timeline and cost to market, if they're not releasing in September, which we're pretty sure they aren't that leaves us very close to the end of the year. All reports I've seen show 20nm ramp up in the latter half of Q1 beginning of Q2 which puts AMDs refresh 3-6 months prior to 20nm availability.

Granted we expect Nvidia to be late as always, but what advantage does AMD gain going after the 780? They have a good product stack that was tarnished by poor driver support. They're competitive from $400 on down, what good is the $400+ market for AMD from a financial standpoint?

How do they benefit from spending money refreshing 28nm now, only to short it to hit 20nm, or be lane on that node?

Firstly 20nm volume production is Q2 2014 with 20nm wafer volume being 2% of total wafer volume. read page 17 and page 18

http://www.tsmc.com/uploadfile/ir/quarterly/2013/1J5NC/C/TSMC 1Q13 Transcript.pdf

add to it the fact that yields have always been difficult at the start of the cycle and AMD might have felt its not sensible to rush to 20nm. Then there is the Apple rumour. if Apple has signed with TSMC for 20nm, Nvidia and AMD can forget any decent volume until late Q3 2014.

So for all intents 20nm is more of a Q3 2014 event . The other issue is both on transistor perf gain and power , 20 nm vs 28nm is not the big jump that 28nm was from 40nm. planar transistors have run out of steam at 20nm and thats the reason Intel went to 22nm FINFET. the foundries are reducing their risk by transitioning to 20nm planar first and a year later are moving to 16/14 FINFET on a 20nm BEOL with 20nm design rules. so do not expect the first gen 20nm products to be massive perf improvements over GK110 / Hawaii.

the decision to go Hawaii on 28nm must have been made in June 2012 with tape-out happening in Nov 2012. Being a close customer of TSMC , AMD must have had enough information on the progress of 20 nm with regards to risk production and its expected volume production ramp in Q2 2014 and TSMC 20nm capacity situation in 2014. so they made a pragmatic decision to go with Hawaii on a mature TSMC 28nm. and here you are commenting without a clue and without much of an idea.

Finally the fact that GTX 780 / GTX 770 launched in late May and given traditional 12 - 15 month lifecycle the GTX 880 can be expected in late June or early July 2014. if AMD replied with their 20nm Pirate Islands flagship in Sep 2014 it makes a lot of sense.

remember GTX 680 came 2.5 months after HD 7970 and did very well. AMD did not make market share gains at all in Q1 2012, because 28nm capacity was a constraint till H2 2012. by that time Nvidia had released their GTX 670 too and destroyed AMD products in sales.

moral of the story : Its not important to be first to a node. You need to make sure there is not too much of a gap and that you have a very well made and competitive product when you launch.

Also I voted for 28nm. its not even a question. anyway the guys who are doubting will know in 2 months.

 

[nemesismk2]

I think 20nm is possible for the AMD 9970 and I look forward to seeing what it can do.

 

[Jaydip]

None, it is going to be on x nm guys while x=28 or 20 or 16 or crap

 

[Jacky60]

It's going to be 20nm definitely and the sooner everyone realises this and just accepts it the better!

 

[Lonyo]

But the question is... Is it beneficial for AMD to attack the 780 given it's high price and medial market allotment this late in a node?

I think AMD is sitting fine, they don't need to contest a $650 product there isn't a ton of money at that bracket in the first place.


Plus you have to consider timeline and cost to market, if they're not releasing in September, which we're pretty sure they aren't that leaves us very close to the end of the year. All reports I've seen show 20nm ramp up in the latter half of Q1 beginning of Q2 which puts AMDs refresh 3-6 months prior to 20nm availability.


Granted we expect Nvidia to be late as always, but what advantage does AMD gain going after the 780? They have a good product stack that was tarnished by poor driver support. They're competitive from $400 on down, what good is the $400+ market for AMD from a financial standpoint?

How do they benefit from spending money refreshing 28nm now, only to short it to hit 20nm, or be lane on that node?

You need to remember that the GTX780 GPU is used in more than just the GTX780.

If they make a new GPU which has 50% more transistors, to make it the same die size as the Titan, then that gives them a GPU they can use in the high end desktop market, but they can also use it in the HPC market as well.

Plus if 20nm products are going to be limited for quite a while, why not release an entire refresh of all your products, and continue using the lower end 28nm products to accompany your 20nm products for the next 12 months, until you eventually roll out 20nm lower end GPUs once production has ramped up from TSMC?

There is a lot of history in how GPU launches work, and how GPUs are used. Many people in this thread seems to be ignoring or forgetting this fact, and assuming that AMD is going to release a new high end GPU at 28nm which will be replaced in 6 months with a 20nm GPU which is way faster.

If 20nm does come with a new architecture, or tweaked one, a high end GPU on current architecture just with more functional units thrown at it might be more viable for getting sales in the short term with a mature driver and hardware stack than a brand new untested 20nm product.

There are many many reasons it makes sense to release a new lineup of 28nm products now, but it all depends on what exactly they do end up doing. They could equally just release a larger GPU which is a bit faster, costs a lot more, and gets replaced in 6~9 months after launch, who knows.

 

[Zanovar]

It's going to be 20nm definitely and the sooner everyone realises this and just accepts it the better!

This is the talk i like to hear.haha

 

[NostaSeronx]

Evergreen -> 40nm
Northern Islands -> 40/32nm
Southern Islands -> 28nm
Sea Islands -> 28/22nm
Volcanic Islands -> 20nm
Pirate Islands -> 20/14nm

 

[RussianSensation]

AMD doesn't have a ton of resources to toss around, they're not the 800 pound gorilla in the room. They need to play it smart, and a 28nm refresh a few months out from 20nm ramp up isn't a smart move imo.

1. 20nm node is not a few months away. Why this myth keeps being perpetuated on our forums? Where does it say anywhere on AMD/NV's roadmap that we'll get 20nm GPUs before June-July 2014? Please link. If volume production at TSMC is only scheduled for Q1 2014, by the time the inventory is built, we are in Q2 2014.

2. Holiday 2013 sales are a big deal. I don't know how you guys operate but I play the least amount of games from May to September. Most of my gaming happens in the cold months of the year. Guess what from October to April it is cold weather season in US/Canada and most of Europe/Russia, parts of China. These are major gaming markets.

3. BF4 is going to be one of the biggest games this holiday season for PC gaming. Launching your new cards around that game, especially if they outperform the competitor at nearly every price level could be a major boost for your sales.

4. Right now a barely faster 770 sells for $390-450, while AMD's barely slower 7970 1Ghz/GE is going for $290-350. This is a good time to increase ASP by launching 9950 for $399 and taking the 770 out. Suddenly AMD raises ASP by $50-100 and it barely costs them more to make this chip. Even though 7970GE is nearly as good as the 770, because the 770 has a "newer" card allure, it will outsell the 7970GE every time. AMD can't let that happen. Same with 7950 V2 vs. 760.

5. 450mm2 20nm chip will cost a lot more than a 450mm2 28nm chip. Until 20nm matures and wafer demand subsides, AMD can coast for another 8-9 months with refreshed 28nm parts. These parts are not necessarily intended for 7970/GE users either. There are plenty of PC gamers rocking GTX470/560/560Ti/570 and 5850/5870/6870/6950/6970 who have not upgraded yet. The 780 impressed them but maybe over the summer they spent $ on house renovation, landscaping, travel, golf/sports equipment, etc. Once October-November rolls around, they might start looking at their GPU upgrade to play games over the long winter.

Nvidia released GT200 on 65nm when 55nm was 100% ready to go, so it's not unprecedented for a company to wring out as many products as possible on a proven node process before moving on.

Furthermore, after 7970 came out, RR made several statements that going down to lower nodes as quickly as possible is no longer economical. His new direction for AMD is to stay on the same node for as long as possible until the next node is financially viable. This is a change of strategy for AMD. 9970 launching on 20nm goes counter against everything the CEO has been saying for the last 1.5 years. If you look at AMD as a whole, they have slowed down transitions to lower nodes on their CPU/APU side as well. Additionally, if 20nm was physically and financially viable for 1st week of October, NV would not be dumb enough to launch GTX780 this summer on 28nm when 1 quarter away you have 20nm.

TSMC's own roadmap discusses risk production for 20nm SoC chips, not high performance computing. People need to be able to separate various types of 20nm transistors intended for different markets too.

Evergreen -> 40nm
Northern Islands -> 40/32nm
Southern Islands -> 28nm
Sea Islands -> 28/22nm
Volcanic Islands -> 20nm
Pirate Islands -> 20/14nm

The bolded parts never happened which makes your entire road-map after Sea Islands mobile parts a pure guess.

I think AMD can get 780 performance (or ever so slightly better) out of a 385-400mm^2 GPU and sell it for $550 and make just as much money per card as Nvidia does off the gtx780.

That's not realistic, unless they drop DP. 7970 is a 365mm2 chip. You can't only add 35mm2 and suddenly match 780s unless your chip runs at 1400mhz. I think 430-440mm2 is what they'd need bare minimum. You are also giving AMD a lot more credit than they deserve. You think NV needs to make a chip 40% larger to match AMD's 400mm2 in performance? Do you think NV put placebo transistors inside their GK110?

I don't think this will be as much of a change from Tahiti as Cayman was from Cypress. At 40nm AMD released 2 different GPU families with significant architecture overhauls.

You must be talking about a different GPU series that never got released. Cayman and Cypress are exact same GPU architecture. There is almost no efficiency gained going with Cayman. It's peanuts. All they did was remove a redundant unit to move it from VLIW-5 to VLIW-4. The major difference was doubling of VRAM and upgraded geometry engine. If you look at 6970 vs. 5870 in non-tessellated & < 1GB of VRAM games, the performance is usually very close.

6970 vs. 5870 - last 3 games tested at GameGPU
+2 fps in Castlevania
+3 fps in Divinity Dragon Commander
- 1 fps in Payday 2

If AMD doubles ACEs, increases geometry engines to 3-4 and goes from 32 to 40-48 ROPs, that would be a monumental change compared to what they did from 5870 to 6970.

Keep in mind the more transistors Nvidia and AMD pack into their GPU's, the more difficult these GPU's are to make which may also be another reason why so many people feel like these GPU's are coming out "late."

So far not a single person on our forum has stepped up and mathematically disproved my comment that the move from any GTX580 OC to Galaxy HOF 780 OC is the greatest move NV did in 1 generation since 7900GTX to 8800GTX. It seems because Haswell flopped large, everyone is bored nowadays and they have $ to burn in their pockets. They want 20nm asap because the money that they saved up to spend on Haswell is now just money wasted for 95% of PC gamers rocking i5 2500-2600-2700-3570-3770, etc. I blame Intel for the disgruntled PC gamer. This generation's GPU increases are very good, especially with 780 OC. I bet it just doesn't sit well with many PC gamers with $600-700 sitting around for a GPU upgrade to drop it on a nearly 2 year old 28nm GPU tech.

 

[BallaTheFeared]

-snip-

We'll see what happens. :thumbsup:

 

[_Rick_]

I voted 20 nm possible, because there's no point in AMD releasing a 28 nm product at this stage. 20nm isn't that disastrously behind schedule, and while the HD 7 line looks old, a pure marketing refresh like nVidia did could have taken care of that. There's no point in getting a new architecture on an old node out, only a few months before you can launch a product on a new process.

From a marketing point of view, the best thing would be to announce a 20 nm product, that will be somewhat available in November, with review samples out to the press a good time before that. Build the hype, and prevent people from buying nVidia, by getting sucked in. It's not like they'll lose money by not selling old HD7 inventory, due to an announcement.

The only alternative would be by pulling something similar to the CPU-section, and dumping a scaled up 7970 onto the market, to stave off Titan. Kill some of the non-gaming capabilities, and go for a big die, and you may attract some of the "I wish I could get single-GPU performance like my current SLI setup" people, as well as the "look at my faster GPU" people. That market is slow, but so far HD7 should do well as a value proposition. It's only the slightly lower power efficiency, due to the more GPGPU oriented architecture that really lets them down in the $200-300 market.
ROI on such a halo product would be a big sucker, unless they also do a badge refresh on lower end cards, to have some of the halo rub off on the lower end cards.

 

[jpiniero]

I have to agree with RS. A new architecture is possible, but it's going to be 28 nm.

 

[PPB]

That's not realistic, unless they drop DP. 7970 is a 365mm2 chip. You can't only add 35mm2 and suddenly match 780s unless your chip runs at 1400mhz. I think 430-440mm2 is what they'd need bare minimum. You are also giving AMD a lot more credit than they deserve. You think NV needs to make a chip 40% larger to match AMD's 400mm2 in performance? Do you think NV put placebo transistors inside their GK110?

No, but Tahiti has lots of redundancies that at this point can be stripped because of the maturity of 28nm and you can still get good yields without them (as opposed to where tahiti launched when 28nm was still in it's early stages).

 

[ShintaiDK]

No, but Tahiti has lots of redundancies that at this point can be stripped because of the maturity of 28nm and you can still get good yields without them (as opposed to where tahiti launched when 28nm was still in it's early stages).

You would unlock redundancies. Not strip them if yield got better.

You simply disable clusters/memory controller ports to bin it differently in terms of defects. And enable more and upbin a product if yield improves.

AMD already did what they could by upping frequency.

The only solution is to make an even bigger chip and/or cut the compute power in favour of gaming power.