[Xbit] TSMC Has Started Volume 20nm Production Ahead of Schedule

[Genx87]

I doubt they can get silicon out and packaged by the end of march. At least in volume enough to be worth talking about.

 

[OCGuy]

I doubt they can get silicon out and packaged by the end of march. At least in volume enough to be worth talking about.

Anytime in H1 for 20nm GMXXX would be impressive.

 

[blackened23]

TPU posted an article this morning on GM117:

http://www.techpowerup.com/196956/nvidia-readies-geforce-gtx-750-ti-based-on-maxwell.html

NVIDIA's next-generation GPU architecture, codenamed "Maxwell," will debut this February, with the unexpectedly positioned GeForce GTX 750. The card will launch on February 18, to be specific. Maxwell will introduce a host of new features for NVIDIA, beginning with Unified Virtual Memory. The feature lets the GPU and CPU share the same memory. Such a feature is already implemented on the current CUDA, but Maxwell could be designed to reduce overhead involved in getting the thing to work. The next big feature is that Maxwell GPUs will embed a 64-bit ARM CPU core based on NVIDIA's "Project Denver." This CPU core will allow the GPU to reduce dependency on the system's main processor in certain GPGPU scenarios. Pole-vaulting the CPU's authority in certain scenarios could work to improve performance

Getting back to the GeForce GTX 750 Ti, NVIDIA's aim is simple, to see how "Maxwell" performs on the existing, proven 28 nanometer silicon fab process, before scaling it up on the future 20 nm nodes, with bigger chips. Given its name, we expect it to be positioned in between the GTX 760 and the GTX 660 in terms of gaming performance, but we won't be surprised if it falls into an entirely different league with GPGPU. There are no specifications at hand.

 

[Nintendesert]

Why would Nvidia test Maxwell on 28nm if TSMC's 20nm was ahead of schedule? NVIDIA expecting yields to be so bad Maxwell may be put entirely on 28nm?

I want Maxwell on 20nm. I don't think NVIDIA is hedging their bets with 28nm Maxwell if they had as much faith as this forum does on TSMC being on schedule with volume and yields. Which to me is very disappointing.

 

[OCGuy]

Why would Nvidia test Maxwell on 28nm if TSMC's 20nm was ahead of schedule? NVIDIA expecting yields to be so bad Maxwell may be put entirely on 28nm?

I want Maxwell on 20nm. I don't think NVIDIA is hedging their bets with 28nm Maxwell if they had as much faith as this forum does on TSMC being on schedule with volume and yields. Which to me is very disappointing.

You realize "Hawaii" was put on 28nm when it was a 20nm part from the start? It is very common.

 

[Rezist]

You realize "Hawaii" was put on 28nm when it was a 20nm part from the start? It is very common.

Which makes it seem like both nVidia and AMD didn't think 20nm would be here all that soon. I would be surprised if either get any 20nm volume compared to apple/qualcomm who probably will be willing to pay a fair penny more.

 

[ShintaiDK]

So, seems those Q1 Maxwell predictions may have been right all along...

It takes 3 months from production start till you see the first chips. So Q1 is wrong. 20nm products is a Q2 event. And NVidia and AMD is most likely last in the line, after Qualcomm and Apple. Plus the process is not targetted at GPU performance so to say like 28nm.

If lucky we see new 20nm GPUs in Q3. Most likely later.

 

[Grooveriding]

If lucky we see new 20nm GPUs in Q3. Most likely later.

^^^^^^^^

More importantly, isn't this initial TSMC 20nm ramp their 'crap' 20nm, and the real deal with Finfet is where the gains are going to come for GPU type chip production ?

 

[Rvenger]

Maxwell should feature DX 11.2 support on the hardware level right? My reference 290 in the summer months will probably get too hot unless I switch to water. I am also getting tired of the stuttering from the 13.12 driver and Windows 8.1.

 

[Slomo4shO]

Maxwell should feature DX 11.2 support on the hardware level right? My reference 290 in the summer months will probably get too hot unless I switch to water. I am also getting tired of the stuttering from the 13.12 driver and Windows 8.1.

You should consider picking up an after market cooler for your GPU. It is also a good time to sell the card with profits :whiste:

 

[Cloudfire777]

Lets do this again guys

28nm reached volume production, October 24th 2011
We got GK104, GK106 and GK107 in late March 2012. 5 months after 28nm volume production started.


20nm reached volume production, January 16th 2014
January > February > March > April > May > June
That is 5 months.

But according to TSMC, they expect 20nm production to ramp up 30% faster than 28nm.

Design ecosystem on 20nm has been validated in real products and is ready to support customers. Yield learning is in line or better than the 28nm path. We expect a fast ramp of 20nm next year, with revenue from 20nm in 2014 bigger than that of 28nm in 2012. You see 20nm will be starting next year whereas 28nm actually started in the fourth quarter – third, fourth quarter of 2011. So the corresponding point for 28nm was 2012. But our ramp in 20nm in 2014 is going to be faster than the ramp for 28nm in 2012. While our 28nm ramp was a record for TSMC, 20nm ramp will be even faster by about 30%,” said Mr. Chang.

Meaning we might see our first 20nm GPUs BEFORE June. April/May perhaps?

 

[ShintaiDK]

Faster is not reference to time, but to volume.

 

[Cloudfire777]

Yes but volume is obviously crucial for the clients to be able to push out products based on it.
If Nvidia can harvest more wafers quicker than 28nm, we could see finished GPUs out faster as well.
If yield is better on 28nm, which was catastrophic for Nvidia, it gets better.

TSMC are already volume producing 20nm on the two biggest fabs they have

 

[Arkadrel]

Would be neat if there really is 20nm stuff out by April.

More importantly, isn't this initial TSMC 20nm ramp their 'crap' 20nm, and the real deal with Finfet is where the gains are going to come for GPU type chip production ?

FinFETs do want!

But that wasnt on 20nm for TSMC was it?

 

[ShintaiDK]

Yes but volume is obviously crucial for the clients to be able to push out products based on it.
If Nvidia can harvest more wafers quicker than 28nm, we could see finished GPUs out faster as well.

TSMC are already volume producing 20nm on the two biggest fabs they have

It takes 3 months from start to end. And then you have the issue of price per wafer. Apple and Qualcomm is willing to pay much more than AMD and nVidia.

And AMD and nVidia already rebranded the mobile segment again. Plus released new desktop cards.

 

[ShintaiDK]

Would be neat if there really is 20nm stuff out by April.




FinFETs do want!

But that wasnt on 20nm for TSMC was it?

TSMC and GloFo calls their 20nm with Finfets for 14 and 16nm. Pure marketing BS. And 14nm with finfets is already being called 10nm.

Also shows on this graph:

 

[Cloudfire777]

It takes 3 months from start to end. And then you have the issue of price per wafer. Apple and Qualcomm is willing to pay much more than AMD and nVidia.

And AMD and nVidia already rebranded the mobile segment again. Plus released new desktop cards.

I see Apple mentioned everywhere 20nm is discussed. They won`t be any danger to Nvidia or the other partners. Which have been established partners at TSMC for a long time before Apple stepped in. I see mention on forums that Apple is occupying 90% of 20nm production, which is BS

Which this statement explains:

The company is expected to ship 165,000 20nm chips to Apple next year, accounting for 10 percent of its revenue that year, Daiwa Capital Markets analyst Eric Chen (陳慧&#26126 said.

http://www.taipeitimes.com/News/biz/archives/2013/12/13/2003578882

Xilinx is already shipping 20nm SOC`s made from TSMC
QCOM too.

We could actually have 20nm GPUs here before June if everything goes right

 

[ShintaiDK]

TSMC isnt a charity company. Whoever pays most gets the wafers. While it may be 10% of the entire year. It will be a lot more of the initial production.

Your link shows no such thing as you claim. Simply that they taped out 20nm chips. And we had no news of AMD or nVidia having successfully taped out of 20nm chips. Meaning they are far away.

To give you an example, 28nm Kepler was taped out in summer 2011. 9 months before its release. Tahiti was taped out in Q2. Released in early 2012. Also showing a rough 9 months from tape out to product.

We havent heard anything about nVidia or AMD and tape out of 20nm parts either.

Your link however says that 20nm will first contribute to revenue in Q2.

 

[Cloudfire777]

Taped out? Yes an article from Fudzilla which is as reliable as SemiAccurate`s articles... Heh.

10% of the whole 2014 means they are tiny partner compared to the biggest ones at TSMC. Stop twisting the words. It is what it is. That is not even nearly close to 90% of the 20nm production as some people mention when they think its all doom. Apple most likely will be producing chips at Samsung as well as TSMC, while Nvidia rely completely on TSMC which is why they might be bigger.
"It will be a lot more of the initial production"
Care to back that up with a link?

Let me repost this since you clearly didnt read it

Lets do this again guys

28nm reached volume production, October 24th 2011
We got GK104, GK106 and GK107 in late March 2012. 5 months after 28nm volume production started.


20nm reached volume production, January 16th 2014
January > February > March > April > May > June
That is 5 months.

But according to TSMC, they expect 20nm production to ramp up 30% faster than 28nm.

 

[Imouto]

You posted the same link 4 times already in 2 pages. Isn't that a bit overkill? You made your point.

Anyway Nvidia is a Samsung customer too, just sayin'.

 

[Lonyo]

Why would Nvidia test Maxwell on 28nm if TSMC's 20nm was ahead of schedule? NVIDIA expecting yields to be so bad Maxwell may be put entirely on 28nm?

I want Maxwell on 20nm. I don't think NVIDIA is hedging their bets with 28nm Maxwell if they had as much faith as this forum does on TSMC being on schedule with volume and yields. Which to me is very disappointing.

28nm non high end Maxwell makes perfect sense.
Yield are higher, chips aren't really more expensive due to 20nm having higher wafer costs, and you aren't going to be temp/power constrained due to being a lower end card.

Early chips on a brand new process can be more expensive than the same design on an older process (e.g. 400m transistor chip on 28nm could be cheaper than 400m transistor on 20nm even though the 20nm chip is smaller, because the wafer is more expensive)
The main advantage is that eventually the new process might be cheaper, and you can push harder at the high end because you reduce your power/size constraints due to scaling.

That's why a lot of companies (as in, other industries) are happy making chips on older processes, because a more advanced process doesn't really give much advantage or have a cost benefit.

Straight from the horses mouth:
http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

That's not saying NV don't want 20nm, they do, but it's not always necessarily beneficial to have a lower end early process 20nm product over a very mature 28nm product, since the 28nm can be cheaper.
And yes, the link is 2 years old, but the principle doesn't change.

 

[BrightCandle]

While the transistor cost doesn't go down, density and hence the capability of your chips do improve with a new process. Its not just about cost, especially for a company that has pushed the lithography of the previous process to its limit and physically can't make a bigger chip. Admittedly for the mid to low end it doesn't make sense, but that is why historically the bigger and higher capability chips came first.

 

[ShintaiDK]

Nothing points to 28nm Maxwell.

 

[OCGuy]

TSMC isnt a charity company. Whoever pays most gets the wafers. While it may be 10% of the entire year. It will be a lot more of the initial production.

Your link shows no such thing as you claim. Simply that they taped out 20nm chips. And we had no news of AMD or nVidia having successfully taped out of 20nm chips. Meaning they are far away.

Yes we do. Nvidia has had working Maxwell samples for quite some time.

 

[KompuKare]

Yes we do. Nvidia has had working Maxwell samples for quite some time.

Isn't the another thread here in VC&G about 28nm 750 Ti (GM107?):
working Maxwell =! working 20nm Maxwell.

Just saying. Testing a new design on the old node makes good risk mitigation sense (Intel's Tick-Tock approach) and it seems Nivdia learned their lessons from Fermi.