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

[ShintaiDK]

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.

That 750 ti is a kepler.

 

[ShintaiDK]

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

I havent seen any.Could you link me?

 

[KaRLiToS]

I will not buy 20nm until second generation of cards so it can mature a little.

(I always promise myself not to buy the next gen but I always end up buying it on release week)

 

[StrangerGuy]

We are about to hit 20nm for GPUs, gone through multiple designs over the past 6 years and yet we still don't have games that make Crysis look dated. If anything IQ to hardware power has got worse.

Sometimes we really need to take a step backward and look where we are going. And I don't think this is the correct direction.

 

[Cloudfire777]

But seriously guys:

I haven`t paid attention to generations before Fermi, but have Nvidia ever put a new architecture on the same node as previous architecture?

I know Intel does: New node > New architecture > New node > New architecture
But I`m convinced Nvidia and AMD does both at the same time.

Why would they change now?

:hmm:

 

[Imouto]

AMD always releases GPUs on new nodes way before Nvidia and there are no info at all about AMD on the new node.

Why would that change now?

I'd get excited about a new Nvidia GPU in a new node after AMD doing so.

 

[Lonyo]

But seriously guys:

I haven`t paid attention to generations before Fermi, but have Nvidia ever put a new architecture on the same node as previous architecture?

I know Intel does: New node > New architecture > New node > New architecture
But I`m convinced Nvidia and AMD does both at the same time.

Why would they change now?

:hmm:

They don't. 260/280 was held on the old node (65nm) while AMD released at the same time and went with 55nm.
They also put low end 200 series chips onto 40nm before Fermi was released...
AMD did the 4770 was a 40nm before their new line was put onto it as well.
So no, you are wrong. It's only with Kepler than the node and new family coincided really.

 

[Cloudfire777]

I didn`t mean that AMD and Nvidia release their cards at the same time. I meant that they both do the new architecture+new node when going toward a new architecture.

According to Wikipedia:

300 series:

• 486M 40nm (GT216)
• 727M 40nm (GT215)
• 754M 40nm (GT215-301)
• 260M 40nm (GT218)

400 series:

• 585M 40 nm (GF108)
• 1,170M 40 nm (GF106)
• 1,950M 40 nm (GF104)
• 1,950M 40 nm (GF114)
• 3,200M 40 nm (GF100)

500 series:

• 585M 40 nm (GF108)
• 1,170M 40 nm (GF116)
• 1,950M 40 nm (GF114)
• 3,000M 40 nm (GF110)
• 292M 40 nm (GF119)

600 series:

• 585M 40 nm (GF108)
• 1,170M 40 nm (GF116)
• 1,950M 40 nm (GF114)
• 1,270M 28 nm (GK107)
• ???M 28 nm (GK208)
• 2,540M 28 nm (GK106)
• 3,540M 28 nm (GK104)
• 292M 40 nm (GF119)

Fermi had the same node as the GT series. They didn`t do any jump there until Kepler. Which obviously wont be the case here with Maxwell since 20nm is coming this year.
So we do know that they have done a new architecture on the same node as previous architecture, but have never mixed different nodes on the same architecture.

 

[Lonyo]

My point was that NV chose to note use the latest node when they released the 260/280, even though it was available.
Also listing the node something is made on has no bearing on when that node was first used.

The first products for 40nm were low end chips from both AMD and NV.
NV have released a new architecture on an old node even though a new one was available.
You took one data point, 28nm, and assumed it to be representative for what might happen at 20nm.

 

[boxleitnerb]

28nm reached volume production, October 24th 2011
20nm reached volume production, January 16th 2014

Nvidia`s 28nm Kepler saw its first light in late March 2012. 5 months after 28nm mass production started.
Yields were horrible on 28nm for Nvidia.

20nm production will be 30% faster than 28nm production according to TSMC.

GK107 was the first Kepler and it was shipped to OEMs in December. Source for yields? Aside from Charlies bashing, that is...

 

[OCGuy]

AMD always releases GPUs on new nodes way before Nvidia and there are no info at all about AMD on the new node.

Why would that change now?

H1 tape-out for AMD is the latest rumor. I'd say people should be excited about both, and you can't really apply previous generations to the future.

The TSMC made 20nm planar process technology is supposed to go online in February 2014 so AMD is expected to tape out their first 20nm based product in 1H 2014. The new 20nm technology is supposedly going to featured in AMD’s next generation graphic processing units which would replace the recently released AMD Volcanic Islands parts. From previous information, we know that the 20nm based lineup would supposedly be known as the AMD Pirate Islands series.

http://www.technationnews.com/2013/10/21/amd-tapes-out-20nm-and-14nm-finfet-chips-in-1h-2014/

I fully expect Nvidia's 20nm to beat AMD's to the market, however.

 

[boxleitnerb]

AMD beat Nvidia on 40nm and tied them on 28nm (see OEM shipments. Just because you didn't see a product in the retail channel, doesn't mean it doesn't exist). Before that (i.e. before 2009) it was pretty even iirc. It certainly isn't a general rule that AMD is faster with new processes.

 

[kapulek]

H1 tape-out for AMD is the latest rumor. I'd say people should be excited about both, and you can't really apply previous generations to the future.



http://www.technationnews.com/2013/10/21/amd-tapes-out-20nm-and-14nm-finfet-chips-in-1h-2014/

I fully expect Nvidia's 20nm to beat AMD's to the market, however.

I expect SoCs on TSMC's 20nm process from Apple and other ARM licencees, maybe AMD's next low power x86 core.
16nm APUs/GPUs late Q4/early 2015. 14nm for APUs/CPUs H1 2015.

 

[moonbogg]

I'd expect decent 20nm GPUs to actually be purchasable around June at the earliest.

 

[Jovec]

We are about to hit 20nm for GPUs, gone through multiple designs over the past 6 years and yet we still don't have games that make Crysis look dated.

These things are unrelated.

 

[Imouto]

AMD beat Nvidia on 40nm and tied them on 28nm (see OEM shipments. Just because you didn't see a product in the retail channel, doesn't mean it doesn't exist). Before that (i.e. before 2009) it was pretty even iirc. It certainly isn't a general rule that AMD is faster with new processes.

At 55nm AMD was 6 months earlier than Nv i.e. HD 3870 and GTX 9800GTX+.

I generally accept something is done in a node when you can purchase it since it's the only thing that you can actually prove.

 

[Mondozei]

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.

On the other hand, Apple won't need their SoC's until Q3. Qualcomm's S805 is on 28 nm anyway and it will take time for a new chip on 20 nm, sp probably H2.

Maxwell's desktop launch for the 800 series(not the 750 Ti) is end of march, that's just a week or so away from Q2.

We could well see 20 nm GPU's from Nvidia in May. They've done paper launches before.
But are there reasons to be skeptical? Sure. But the case for doom is oversold.

 

[Mondozei]

That 750 ti is a kepler.

Keplers have 64 bit Denver CPU's? Who knew!

We are about to hit 20nm for GPUs, gone through multiple designs over the past 6 years and yet we still don't have games that make Crysis look dated. If anything IQ to hardware power has got worse.

Sometimes we really need to take a step backward and look where we are going. And I don't think this is the correct direction.

This is the correct direction. It's just going slower than many of us want it to go.
As for the performance in games etc. Look at BF4 vs BF3. I wouldn't say there's a massive difference in graphics, but BF4 is also a lot more demanding, even after the patches.
The changes are more subtle right now, but they give definitive impact on our systems. It's a phase until the next leap.
Games are slowly starting to take a more photorealistic direction(key word here is direction, actual photorealism is probably at least a decade away).

This game looks a lot better than Crysis 3 will ever look.
By a lot.

https://www.youtube.com/watch?v=EWJD8DMasCU

 

[Cloudfire777]

GK107 was the first Kepler and it was shipped to OEMs in December. Source for yields? Aside from Charlies bashing, that is...

GK107 wasn`t officially announced until March 22nd. GTX 680 (GK104) was announced the day after.

Source for yield problems?
There is a ton of them out there. Here is one:
http://www.xbitlabs.com/news/graphi...er_Than_Expected_Chief_Executive_Officer.html

Hopefully they have learned from that experience and applied techniques that helps them on 20nm production.

I'd expect decent 20nm GPUs to actually be purchasable around June at the earliest.

I think we will have 20nm GPUs available somewhere between April-June.

 

[boxleitnerb]

GK107 wasn`t officially announced until March 22nd. GTX 680 (GK104) was announced the day after.

So? GK107-based notebooks needed no announcement to exist (and ship). There is no correlation between the existence of a GPU and its announcement, none whatsoever.

Source for yield problems?
There is a ton of them out there. Here is one:
http://www.xbitlabs.com/news/graphi...er_Than_Expected_Chief_Executive_Officer.html

That says nothing. Lower than expected means nothing without knowing what percentage they were expecting (process yields and binnig yields being different, that makes things even more complicated). It also says nothing about yields of other companies in comparison. Without internal knowledge we will never know how exactly the yields are for company A, B or C, no matter what people want to believe.

 

[Cloudfire777]

EDIT:

Meh, I have other things to do.

 

[Enigmoid]

GK107 wasn`t officially announced until March 22nd. GTX 680 (GK104) was announced the day after.

Source for yield problems?
There is a ton of them out there. Here is one:
http://www.xbitlabs.com/news/graphi...er_Than_Expected_Chief_Executive_Officer.html

Hopefully they have learned from that experience and applied techniques that helps them on 20nm production.


I think we will have 20nm GPUs available somewhere between April-June.

Anandtech had the Acer notebook with the 640m reviewed in march. Given its status as an OEM notebook part it would have been shipping to Acer months earlier.

 

[lefty2]

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:

That graph was debunked as a pile of FUD at the TSMC conference call. TSMC's 16-FinFET gives a 15% gate reduction compared to 20nm, however on the graph it's shown as the same density. Intel's process is still better by a small amount, but not 35%.

http://seekingalpha.com/article/194...-2013-results-earnings-call-transcript?page=5

 

[ZGR]

28nm Maxwell will definitely exist for the lower end cards. If we conclude high-end Maxwell to come out Q3 '14, when will AMD's Hawaii successor be? I like the competition

 

[boxleitnerb]

I generally accept something is done in a node when you can purchase it since it's the only thing that you can actually prove.

It doesn't matter what you accept. If you think you can snip with your fingers and a GPU falls in your lap, you are dead wrong. Before the GTX 680 launch there were notebooks with GK107 out there, there are videos with them running BF3. It takes lots of time to integrate GPUs into notebook lineups, thus GPUs have to be finished months before.
And besides, the GK107 GPUs were purchased - by the OEMs, just not by you or me.