Interesting Article: Waiting on 20nm graphics cards from Nvidia and AMD? Don’t bother

[SineDave]

http://techsoda.com/no-20nm-graphics-amd-nvidia/

I thought it was pretty well reasoned out, if a little pessimistic. I guess it means we'll be waiting another 6 months for the new architecture.

 

[TaintedSquirrel]

Much more likely for me is that Pirate Islands – like Maxwell – will be top-to-bottom 28nm

I can't see this happening.
If it were going to happen, it would have happened already. From Nvidia, anyway, we'd know something by now.

I wouldn't be surprised to see a 290/290x refresh from AMD but if Nvidia releases a new 28nm series I'll eat my hat.

 

[Yakk]

Doesn't seem... too ...far off the mark....

 

[SineDave]

I can't see this happening.
If it were going to happen, it would have happened already. From Nvidia, anyway, we'd know something by now.

I wouldn't be surprised to see a 290/290x refresh from AMD but if Nvidia releases a new 28nm series I'll eat my hat.

I can believe him though, because he makes a good point. AMD may be focused on die size increases rather than process at the moment given the issues with 20nm. That's still a fairly significant R&D requirement, yet it may be "cheaper" than pushing through the issues with TSMC/GloFou.

 

[spine]

Seems a very well justified position.

I would expect to see Nvidia release a 28nm 880 part later this year though. It'll be another 'midrange masquerading as high-end' specced part like the 680 was I expect.

Better than a 780ti, but not by much.

2015 will be when the real Maxwells hit, and I don't think there's much doubt about that general position.

 

[SixFootDuo]

This is good news for me as I've already decided to just keep my pair of crossfired 290x's until at least the middle or end of 2015.

19 months ... $65 aside each month = $1,235 ... should take of a pair of new 20mm Nvidia or AMD video cards

 

[veedubfreak]

This is good news for me as I've already decided to just keep my pair of crossfired 290x's until at least the middle or end of 2015.

19 months ... $65 aside each month = $1,235 ... should take of a pair of new 20mm Nvidia or AMD video cards

Same here, but I lucked out and got a third 290 with waterblock for 400 bucks

 

[Liger88]

We've been in the doom and gloom phase since 45nm and it'll just continue to get worse. They've really been finding ways to continue the process since then and unlike times past those tweaks only last for a shrink or two before something else is needed to overcome the barriers.

Regardless of whether Nvidia drops a 20nm Maxwell product anytime soon, we are living in interesting times. I have a feeling the industry is going to go through a sinking phase far quicker than a revolutionary one. If they had some serious contenders to existing technologies it would make little sense trying to milk silicon with the exponential rising costs. Will be curious to see things at the 10nm stage (if Intel can even get there) and below. I have doubt that the rest of the fabs will be able to accomplish those nodes in a timely manner.

 

[Zomoa]

I could see them going with a "big die" solution to bridge the gap.

They've made serious advances in the power consumption department lately. They can use this headroom .

 

[SineDave]

Makes me a little more comfy with the current gen investment anyway

 

[Sky15]

Pascal or bust for me.

 

[euskalzabe]

Pascal or bust for me.

Agreed! I was excited for Maxwell but seeing that it would take so long I ended up getting a 770 for cheap and I'm now set for a while. I can probably hold off until Pascal now, it's only 2 generations away and if my 770 starts showing age I can always pair it with a 2nd one for an even cheaper price in 2015!

 

[Relayer]

This is good news for me as I've already decided to just keep my pair of crossfired 290x's until at least the middle or end of 2015.

19 months ... $65 aside each month = $1,235 ... should take of a pair of new 20mm midrange Nvidia or AMD video cards

Fixed that for you.

 

[SineDave]

I haven't done the math, is the 770 able to handle QHD at max settings for current gen games?

 

[Liger88]

I haven't done the math, is the 770 able to handle QHD at max settings for current gen games?

Pfffttt! Excuse me while I put my drink back in my mouth.

Honestly, that depends on what your tolerated frame rate is. I can run 4K on a 660 max settings if I wanted to.

 

[bigdogchris]

I didn't know people care what the manufacturing process is when they buy a GPU.

 

[The Mac]

I dont think people care so much as enjoy speculating on the the improvements from node to node.

 

[TraptPatriot]

I read the entire thing, and it was a very interesting read. Most of it made sense, but without direct contact with the companies and their confirmations, it is all still just speculation. I started to wait for the 8xx series and then decided it wasn't worth it, can't even get a definitive timeframe. Picked up the 780ti instead, and I am sure it will still do well versus the eventual 8xx series.

 

[SineDave]

Pfffttt! Excuse me while I put my drink back in my mouth.

Honestly, that depends on what your tolerated frame rate is. I can run 4K on a 660 max settings if I wanted to.

Touche. Suppose "good frame rates" is a prerequisite.

 

[172floater]

I guess Nvidia is going to milk the 700 series even more this year from the sound of that, or just launch rebranded 700 series as the new 800s

 

[Liger88]

I guess Nvidia is going to milk the 700 series even more this year from the sound of that, or just launch rebranded 700 series as the new 800s

They can't really re-brand the 700 series anymore than they already have unless they really start pushing GK110 down to the lower tiers. The difference between just between 600-700 Series cards in the same category is almost a joke (maybe biggest push was the 680 to 770 price wise).

With the 30% power savings of the Maxwell architecture it could still provide a decent pre-Maxwell release. Higher SP density, but a bigger die due to the huge increase in L2 cache. Would be some big ass chips on 28nm, but still a good boost, just not great.

Maxwell is the most hyped Nvidia product (5 years publicly) and they've stated they don't want to do it half-assed because the technology (TSMC) isn't ready. Chances are they are waiting as long as they have to, but it would be interesting to see 28nm Maxwell with a rapid turn around on the 900 Series (9 months or less) which would piss a lot of people off. Or they could do 28nm on the lower tiers with 20nm limited high-end parts until the 900 Series. Chances are that Nvidia is going to play it safe and wait for 20nm.

The delay just means it'll give 16nm more time to mature. Once 20nm hits we'll have about 4 years before we face this same debacle.

 

[Spazturtle]

Does Maxwell scale on 28nm though?

The 750ti shows us that on 28nm that Maxwell has a larger die size then Kepler. Maxwell being around 25% larger.

So at the high end the die size would be very large and might not be profitable.

At the moment AMD are winning on die size (same performance at smaller die size).

The 290x is 438 mm² whilst the 780ti is 551 mm²

A high end Maxwell would be around 689 mm² (if the 750ti is any indication) which is just not going to happen.


At this rate Samsung/GloFou will have their 20/14nm (20nm die with 14nm transistors) fabs online before TSMC gets their 20/20nm (20nm die and transistors) fabs online.

 

[TaintedSquirrel]

They can't really re-brand the 700 series anymore than they already have unless they really start pushing GK110 down to the lower tiers. The difference between just between 600-700 Series cards in the same category is almost a joke (maybe biggest push was the 680 to 770 price wise).

780 Ti becomes the GTX 880, 780 becomes the GTX 870. Done. Give them both a small factory overclock, maybe 10%, and call it a day. We know both the 780 and 780 Ti are good for it.

If it comes with a price-cut then maybe they'd do it. Maybe get the "780" down to $399, the "Ti" down to $549.
I think it would be embarrassing for them, maybe even admitting defeat as far as development goes especially if AMD has something up their sleeves (of course they don't, who am I kidding).

They have bigger problems in the mid-range since they can't possibly rebrand the GTX 680 again.

Here's my question, do AMD/Nvidia have the resources to develop TWO GPU product series at the same time? Those being another 28nm series, as well as Maxwell/Pirate Islands 20nm cards. If they were really planning another line of 28nm cards they'd have to be developing them alongside their new stuff.

 

[hcforde]

What games are coming where there would be a significant need for the benefits of a 20nm process? The only tech we can use that can really benefit is 4K in my opinion and we have not seen the new 60FPS/HZ monitors arrive yet. I think we are really good for a while with out the 20nm chips being dangled in front of us.

 

[FaRKle0079]

Wow there sure was a lot of conjecture in that article...

 

[LordEC911]

At this rate Samsung/GloFou will have their 20/14nm (20nm die with 14nm transistors) fabs online before TSMC gets their 20/20nm (20nm die and transistors) fabs online.

Might want to read up on the process differences again.
Anyway TSMC already has their 20nm SOC ramping volume production.

 

[Spazturtle]

Might want to read up on the process differences again.
Anyway TSMC already has their 20nm SOC ramping volume production.

Ill believe it when I see it, TSMC are the most unreliable foundry out there. Heck NVIDIA said that TSMCs 20nm fabs where "essentially worthless". Other fabs will likely be on 14nm before TSMC get their 20nm shit together.

 

[172floater]

What games are coming where there would be a significant need for the benefits of a 20nm process? The only tech we can use that can really benefit is 4K in my opinion and we have not seen the new 60FPS/HZ monitors arrive yet. I think we are really good for a while with out the 20nm chips being dangled in front of us.

Well I think with the new Rift dev kits and consumer version getting closer to release which is suppose to have an even higher res screen, I think the new 20nm cards would benefit users greatly who want to max out games with rift support.

 

[octoberasian]

I know CPU and GPU are two entirely differently things in terms of manufacturing and production, how is it then that Intel is able to do 22nm and will soon have 14nm processors with relative ease? Yet, TSMC is having issues providing its two biggest customers-- Nvidia and AMD-- with a feasible, ready-to-produce 20nm products and on time; and GloFo is a few years behind Intel in getting 14nm and 22nm/20nm products ready.

Is it the size of R&D Intel has-- both financially and physically?

 

[ccityinstaller]

I know CPU and GPU are two entirely differently things in terms of manufacturing and production, how is it then that Intel is able to do 22nm and will soon have 14nm processors with relative ease? Yet, TSMC is having issues providing its two biggest customers-- Nvidia and AMD-- with a feasible, ready-to-produce 20nm products and on time; and GloFo is a few years behind Intel in getting 14nm and 22nm/20nm products ready.

Is it the size of R&D Intel has-- both financially and physically?

Yes and Yes. Also, Apple is TSMC's biggest customer (from what the leaks are showing)..Apple orders more chips then AMD/Nvidia combined.

 

[limitedaccess]

I know CPU and GPU are two entirely differently things in terms of manufacturing and production, how is it then that Intel is able to do 22nm and will soon have 14nm processors with relative ease? Yet, TSMC is having issues providing its two biggest customers-- Nvidia and AMD-- with a feasible, ready-to-produce 20nm products and on time; and GloFo is a few years behind Intel in getting 14nm and 22nm/20nm products ready.

Is it the size of R&D Intel has-- both financially and physically?

Intel has both size and history going for it. Also Intel's lead is a bit misleading since they are not really a client oriented fab (they do fab for some third parties but rather limited), as such there are different technical concerns (difficulties) compared to a company like TSMC who needs to cater to third party clients.

Some fun information by the way. The first product for sale using Intel's 14nm process will not be Broadwell, Cherry Trail, or even an Intel chip. TSMC actually fabs chips for Intel (Intel has said that it expects to still use TSMC for 2-3 years for products related to it's Infineon acquisition back in 2010, 6-7 years total. Switching fabs isn't that simple it seems).

Also the latest leaks/rumors say that TSMC is actually ahead of schedule for Apple A8 production on 20nm.

 

[limitedaccess]

Does Maxwell scale on 28nm though?

The 750ti shows us that on 28nm that Maxwell has a larger die size then Kepler. Maxwell being around 25% larger.

So at the high end the die size would be very large and might not be profitable.

At the moment AMD are winning on die size (same performance at smaller die size).

The 290x is 438 mm² whilst the 780ti is 551 mm²

A high end Maxwell would be around 689 mm² (if the 750ti is any indication) which is just not going to happen.

Maxwell (at least with GM107) is smaller for the same amount of transistors compared to Kepler (43% more transistors but only 25% larger die area than GK107). So it is more transistor dense just like Hawaii (290/290x). It also has a much larger than both 25% and 43% performance gain over GK107, so does better both in terms of performance/die size and performance/transistor.

If you want to do a crude theory craft you could scale up the GTX 750ti by 3x at 28nm and you'd end up with a die size of 441mm. 4x would be 588mm (also you could pair it with faster 6/7ghz memory, which is the the largest bottleneck likely for the current 750ti). Hypothetically the numbers show you could scale up and make a faster than GTX 780ti design with a smaller die size and power envelope but just not a huge performance gain. However this would similar in real world terms as the GTX 580->GTX 680 transition.

 

[hcforde]

Well I think with the new Rift dev kits and consumer version getting closer to release which is suppose to have an even higher res screen, I think the new 20nm cards would benefit users greatly who want to max out games with rift support.

1. This is a mass market item, not catering to the uber-small markets like Occulus Rift development
2. Higher res screen than what?
I already mentioned 4K and the limitation of the panels. I really do not think that the first consumer version of the Occulus Rift is going to be 4K but 1080 OR 1440/1600 at best. (Maybe you are thinking that in the Occulus Rift the stated resolution is in BOTH eyes, Nope--fullHD is 960*1080 per eye, 4K would be 1920*2160 per eye) The current 290/290X and 780Ti can do 4K fairly well. If you can afford a 4K monitor you can afford a couple of GPU's(or more) to go with it especially if you are a developer. So not having GPUs manufactured on the 20nm process is not hampering anything. We also need to remember that most games from now on are going to be ports from consoles.

I would never buy high end hardware in anticipation of software. That creates anger & disappointment when content is delayed(any games you know of ever been delayed?). Since content is cheaper than the hardware, purchasing it first is like letting the tail wag the dog. Look at the people that have bought 4K and are now crying because there is no content. I bought 4K, but I bought a cheap one (Seiki 55" for $800) and I have a source for 4K, my computer. I took my computer into Best Buy and showed them what games looked like in 4K (Most of the higher priced 4K panels are being sold for watching movies).

The biggest benefit for the 20nm process will be lower power consumption levels.

 

[octoberasian]

Yes and Yes. Also, Apple is TSMC's biggest customer (from what the leaks are showing)..Apple orders more chips then AMD/Nvidia combined.

Also the latest leaks/rumors say that TSMC is actually ahead of schedule for Apple A8 production on 20nm.

I would assume then that any delay would be related (or partly) to prioritizing the relatively new 20nm production for Apple's new A8 SoC if that is the case?

 

[LordEC911]

Ill believe it when I see it, TSMC are the most unreliable foundry out there. Heck NVIDIA said that TSMCs 20nm fabs where "essentially worthless". Other fabs will likely be on 14nm before TSMC get their 20nm shit together.

It isn't. Nvidia has complained about the last 3nodes... that should give you a hint at where the problem lies.

 

[jwcalla]

Looks like AMD won't be hitting 20 nm this year. No real surprise. As I said way back when those rumors of Apple moving to TSMC were brewing, the GPUs are going to become low priority at TSMC. TSMC was already way over-capacity; taking on a high-volume, super-high priority customer like Apple was just going to force TSMC to shelf other customers. Apple and Qualcomm are likely their top two right now.

AMD is probably wise to work with GF / Samsung to move more production over there. If NVIDIA was smart they'd be making some phone calls to Samsung also. It's too late for this round, but maybe the next one.

 

[LordEC911]

Looks like AMD won't be hitting 20 nm this year. No real surprise. As I said way back when those rumors of Apple moving to TSMC were brewing, the GPUs are going to become low priority at TSMC. TSMC was already way over-capacity; taking on a high-volume, super-high priority customer like Apple was just going to force TSMC to shelf other customers. Apple and Qualcomm are likely their top two right now.

AMD is probably wise to work with GF / Samsung to move more production over there. If NVIDIA was smart they'd be making some phone calls to Samsung also. It's too late for this round, but maybe the next one.

Too late for the next round as well. Does Samsung even have the libraries to fab an ASIC as complex as a GPU?

 

[jwcalla]

Too late for the next round as well. Does Samsung even have the libraries to fab an ASIC as complex as a GPU?

Probably.

 

[xoleras]

I think it should be pointed out that the author of this article is known as SiliconWars on anandtech. Look through several pages of his post history at anandtech. Aside from being a complete AMD fanboy, (which is fine), he also seems to be a complete lunatic. But certainly judge for yourself. View about 5-6 pages of his post history. Very amusing stuff. Certainly I wouldn't take anything from that guy at face value. Complete nut case.

 

[limitedaccess]

I would assume then that any delay would be related (or partly) to prioritizing the relatively new 20nm production for Apple's new A8 SoC if that is the case?

Anything would be speculation obviously without actual inside information.

However it's actually been reported in the past that TSMC had turned down offers from both Apple and Qualcomm (as well as investments) due better guarantee supplies for them. They've actually been on record in past saying they do not want to play favorites with clients being wary of linking their performance with that of their clients.

If you actually read into Nvidia's past comments regarding TSMC their complaint is actually also along these lines, they wanted better (or more special) treatment. Typically Nvidia (among some others) have been the early adopters of new processes and they felt they should be getting more out of that it having to bear the larger initial costs and low yields.

Looks like AMD won't be hitting 20 nm this year. No real surprise. As I said way back when those rumors of Apple moving to TSMC were brewing, the GPUs are going to become low priority at TSMC. TSMC was already way over-capacity; taking on a high-volume, super-high priority customer like Apple was just going to force TSMC to shelf other customers. Apple and Qualcomm are likely their top two right now.

AMD is probably wise to work with GF / Samsung to move more production over there. If NVIDIA was smart they'd be making some phone calls to Samsung also. It's too late for this round, but maybe the next one.

One other thing to take from that article however is they state they are in 20nm development. It doesn't look like AMD will be skipping it as per the OPs article in this thread.

Regarding Qualcomm, it's first 20nm chips are road mapped for 1H2015 and not 2014. It's Q3-Q42014 releases are currently still road mapped as 28nm.

A side regarding AMD, they use TSMC for the PS4 and Xbox One APUs.