How much better should the 20 nm, 800-series be?

[OCGuy]

20nm SoC certainly could be used for a desktop application, sure. But if it doesn't have noticeably better electrical characteristics (for a high performance chip) than 28nm HP, has higher wafer costs to cancel out any savings from smaller dies, why would they bother with it? Especially when 20nm with FinFETs (aka "14nm") is coming so soon afterwards. Not to mention Apple is supposedly buying up as many TSMC 20nm wafers as they can get their hands on, so supplying a full GPU range could well be tricky to begin with.

It's speculation, sure, and I certainly could be wrong. It's just my thinking at the moment.

If I knew for sure I would have my own tech site

20nm could surely be skipped, I am saying there isn't anything that says it will be. They do claim ~30% power or speed, as well as 1.7x density improvements over 28.

There are hints...such as TSMC emphasizing just how quickly 16nm will roll out this year.

On the last TSMC roadmap, 20SoC and 16FinFET both hold the same place in-between High Performance and low power, suggesting we aren't going to be seeing such a wide array of options at each node anymore.

Also NV selected TSMC for 20nm after all the articles about NV proclaiming TSMCs 20nm "worthless." They were given priority for 28nm production and possibly for 20nm as well to keep them on board.

Taipei, Dec. 5, 2012 (CENS)--Industry executives have regarded an Nvidia executive's statement in a recent interview with a South Korean newspaper as the company's inclination to contract Taiwan Semiconductor Manufacturing Co. (TSMC) as its 20nm process foundry provider.

In an interview conducted by Hankook Ilbo, a daily newspaper published by the Hankook Ilbo Media Group in Seoul, Phil Eisler, general manager of GeForce GRID Cloud Gaming at Nvidia, pointed out that Nvidia has approved TSMC's 20nm foundry capacity.

Eisler said although his company has evaluated all foundry options, TSMC remains the best choice for his company. He noted that his company has set up long-term cooperation with TSMC, which has persistently provided Nvidia the optimum portfolio of production capacity, technology and pricing.

Samsung Electronics is also reportedly competing for Nvidia's 20nm foundry contracts.

Samsung's share of world silicon foundry market remains low. According to market research organization Gartner, in 2011 TSMC had 48.8% of the market, far leading United Microelectronics Corp.'s 12.1%, GlobalFoundries Inc.'s 12% and Samsung's 1.6%.

Nvidia, a world leading vendor of graphics processing units (GPUs), has recently branched out into processors for non-Apple tablet PCs and smartphones. The chips require advanced processing technologies.

Industry executives felt that Nvidia's contracts will keep TSMC's advanced processing technologies running at high capacity. They are watching whether TSMC will be able to lure Apple contracts away from Samsung in 2013 after winning Nvidia contracts.

http://news.cens.com/cens/html/en/news/news_inner_42252.html

 

[NTMBK]

If I knew for sure I would have my own tech site

20nm could surely be skipped, I am saying there isn't anything that says it will be. They do claim ~30% power or speed, as well as 1.7x density improvements over 28.

There are hints...such as TSMC emphasizing just how quickly 16nm will roll out this year.

On the last TSMC roadmap, 20SoC and 16FinFET both hold the same place in-between High Performance and low power, suggesting we aren't going to be seeing such a wide array of options at each node anymore.

Also NV selected TSMC for 20nm after all the articles about NV proclaiming TSMCs 20nm "worthless." They were given priority for 28nm production and possibly for 20nm as well to keep them on board.






http://news.cens.com/cens/html/en/news/news_inner_42252.html

I don't doubt that NVidia will release something on 20nm, but I suspect that it won't be a desktop GPU. NVidia is also a vendor of tablet SoCs now, don't forget.

I know that TSMC claims that there will be no differentiation between HP and LP at smaller node sizes, but I'm pretty dubious. Even Intel have got different versions of their 22nm process, one for their HP desktop, laptop and server parts, and one for their Silvermont based SoCs: http://semiaccurate.com/2013/05/10/intel-talks-about-their-22nm-p1271-soc-process-for-silvermont/ Heck, they're actually specialising their processes even more than in the past:

It seems odd that we're getting completely opposite stories from TSMC and Intel. TSMC did originally plan on an LP and a HP planar process, as mentioned in the story I linked above, but then they went down to just one, so I don't buy that it's a choice that they planned out and willingly made.

 

[OCGuy]

It seems odd that we're getting completely opposite stories from TSMC and Intel. TSMC did originally plan on an LP and a HP planar process, as mentioned in the story I linked above, but then they went down to just one, so I don't buy that it's a choice that they planned out and willingly made.

Yea who knows, maybe only Tegra/Mobile on 20nm? Theo did speak to some execs last year who hinted at the merging of HP/LP

First and foremost, TSMC plans to introduce volume products based on 20nm Planar design (CLN20SOC) over the course of 2013, switching to FinFET with the 2014 arrival of CLN16FF...

... All of these processes are high-performance, while the company is keeping its Low Power roadmap mostly hush hush. Truth to be told, we would not be surprised if, very soon, the Low Power and High Performance nodes simply merge, as High Performance nodes are becoming more and more power aware, and just are able to scale higher. At least, that was the general feeling between TSMC executives that we managed to listen to.

http://www.brightsideofnews.com/new...mm-wafers2c-16107nm-finfet2c-cowos--more.aspx



Either way, can't wait for GTC in March, where hopefully we get a little clarity.

 

[tolis626]

Whatever the case may be with the new nodes,I doubt we'll see really significant improvements over high performance 28nm GPUs.Significant as in over 30% improvement.20nm is having problems right out of the gate (LP and HP being the same for TSMC?Smells really fishy),so the maturity of the 28nm node ought to cover some of the disadvantage.I guess someone with a 290/x or 780/ti won't regret their decision any time soon.Sure,by next year everything will be different,but I don't think 2014 will see any significant improvements for GPUs.

On the other hand,new smartphones/tablets will most probably end up very interesting.High-end SoCs are starting to creep into console territory in terms of computing power.While that in itself is meeeh,when you consider that it's a device that's in your pocket all day long and wherever you go,it's pretty darn impressive.Exciting years lie ahead...

 

[norseamd]

IDD clock for clock LMAO its almost as useless as (best) pound for pound (fighter)

pound for pound is one of the worst and most overused ranks ever. as if anderson silva or gsp could ever hope to beat fedor or cain. the most it can do is a comparision of skill or effectiveness in a weight class. but who could beat who no not ever

 

[schmuckley]

Won't overclock past 4ghz due to uncontrollable heat output.
I'd rather have a 45nm chip with 12 real cores and 200w tdp that overclocks and has no CB.

 

[OCGuy]

Whatever the case may be with the new nodes,I doubt we'll see really significant improvements over high performance 28nm GPUs.Significant as in over 30% improvement.20nm is having problems right out of the gate (LP and HP being the same for TSMC?Smells really fishy),so the maturity of the 28nm node ought to cover some of the disadvantage

The 20-30% speed is true, but it would be wrong to assume that is the only place performance will improve.

The increased density capacity will improve performance, as well as architectural changes. We can't be sure just how much the extra chip on Maxwell will help the GPU, but one would have to assume it isn't negligible.

I don't think we can go by Intel's experience @ 22nm, although it has been impressive. Getting 10-15% IPC improvements when the goal has been to lower TDP might not wow CPU enthusiasts, but NV and AMD aren't going to care about TDP at the high-end of the next node.

 

[tolis626]

The 20-30% speed is true, but it would be wrong to assume that is the only place performance will improve.

The increased density capacity will improve performance, as well as architectural changes. We can't be sure just how much the extra chip on Maxwell will help the GPU, but one would have to assume it isn't negligible.

I don't think we can go by Intel's experience @ 22nm, although it has been impressive. Getting 10-15% IPC improvements when the goal has been to lower TDP might not wow CPU enthusiasts, but NV and AMD aren't going to care about TDP at the high-end of the next node.

I don't disagree with you,I just think it'll take long before we see any significant improvements.We will see them at some point though,that's for sure.

With all that said,to be honest I am more curious/excited about what AMD can bring to the table with newer nodes than what NVidia will do.And this isn't my typical "I don't like NVidia" kind of thing.NVidia's current lineup does more than fine with the 28nm process (To the point that I'm considering NVidia for the first time in many years).I expect more of the same for them.You know,better performance,similar efficiency.Nothing that will make us say that the 780ti is too old.On the other hand,I think AMD has a lot more going.They have a chance to give new life to what is,admittedly,a mediocre product.Yeah,Hawaii's performance is great and all,but it consumes way more power (I care about thermals,not power consumption per se) and can't be pushed far.In fact I haven't seen a 290 series GPU go over 1.3GHz stable.At least not without needing a generator feeding it.There is the argument that not all 780/tis can clock too high,but it's far more usual to see a 1.4GHz 780/ti than it is to see a 1.3GHz 290/x.And the latter consumes far more power to boot.So that's my line of thought.AMD can make Hawaii what it could have been,which is a lot more than it is today.

All in all,2014 should be interesting...

 

[ams23]

I don't doubt that NVidia will release something on 20nm, but I suspect that it won't be a desktop GPU. NVidia is also a vendor of tablet SoCs now, don't forget.

Tegra K1 (Cortex A15 and presumably the pin-compatible Denver variant) uses a 28nm HPM fab. process. The upcoming "Parker" SoC in 2015 will use FinFet 3D transistors (presumably with a 16nm fab. process). As for Maxwell on desktop, even though initial variants such as GTX 750 Ti will stick with 28nm fab. process, it is pretty obvious that future variants (with built-in Denver ARM cores and/or much higher CUDA core counts) will use an even more advanced fab. process such as 20nm.

 

[Ajay]

Supposed leak @ videocardz:
http://videocardz.com/49517/nvidia-maxwell-gm107-gpu-pictured-detailed

So, maybe Maxwell will come out on 28nm HPM for mobile and low end desktop. Looking at power consumption, Maxwell will be a big deal - even more so on 20nm. Given the significant increase in xtors (more FMACs, etc.), we should see pretty good jumps in performance even if clocks drop a bit. Then again, this is all guess work till we see real performance an power usage from working Maxwell GPUs.

 

[ams23]

Supposed leak @ videocardz:
http://videocardz.com/49517/nvidia-maxwell-gm107-gpu-pictured-detailed

So, maybe Maxwell will come out on 28nm HPM for mobile and low end desktop. Looking at power consumption, Maxwell will be a big deal - even more so on 20nm. Given the significant increase in xtors (more FMACs, etc.), we should see pretty good jumps in performance even if clocks drop a bit. Then again, this is all guess work till we see real performance an power usage from working Maxwell GPUs.

While videocardz may be correct with various specs, they are wrong about GM107 being "better binned" in order to achieve better energy efficiency. GM107 in GTX 750 Ti has superior energy efficiency compared to GK106 in GTX 650 Ti due to architectural, process, and experiential improvements made over time. GM107 uses a "mobile first" approach, and likely leverages the breakthroughs in energy efficiency made with the Kepler.M GPU in Tegra K1 (which goes to show how important Tegra is to the company for all future products).