NVIDIA's GM204 & GM206 Maxwell Taped Out At 28nm?

H

HardOCP News

[H] News
Joined
Dec 31, 1969
Messages
0
If I am reading this translated news post correctly, rumor has it that NVIDIA's GM204 and GM206 Maxwell will be 28nm instead of the expected 20nm.

Then the two graphics chips are GM204 (high-end segment) and GM206 (performance segment) lie their tape out later this April and have those already possibly are already through (so the rumors timely wise information available are not) . These two graphics chips are provided in the rest yet for 28nm production, not for the 20nm manufacturing (which would require an extra tape-out, respectively, again much preparation time for mass production in doubt).
Click to expand...
 
Yea it's been known for a few weeks that AMD/Nvidia will not be releasing any 20nm products this year.
 
To see intel's fabbing to be hugely superior to anyone else. nvidia (and lol amd) should rent some 22nm or 14nm (soon) time from intel. tsmc and globalfoundries need to step up their game
 
Once we saw 750ti still on 28nm , I assumed we see higher versions(GM204-206) just not the GM210 , top end .
 
Does TSMC even have high-performance 20nm production planned? Or will we have to wait for the 16nm finfet?
 
jwcalla said:
Does TSMC even have high-performance 20nm production planned? Or will we have to wait for the 16nm finfet?
Click to expand...

Something was wrong with TSMC and 20nm. Can't remember where I read up on it though.

Which sucks because I was REALLY looking forward to 20nm high end Maxwell...
 
Who cares really? It seems the real performance comes from the underlying architecture changes in Maxwell, no?

750 Ti has shown what it can do @ 28nm with a 128bit bus. I'll gladly be upgrading to whatever big Maxwell comes out, don't really care about it being on 28nm.
 
Didn't Nvidia say that some of their DX12 features won't make it into the next series of cards awhile back? Think it was when Microsoft showed off Forza running on Nvidia hardware, but I could be mistaken of course. :) I bet they save 20nm tech for the full DX12 card release in 2015.
 
DASHlT said:
Yea it's been known for a few weeks that AMD/Nvidia will not be releasing any 20nm products this year.
Click to expand...

Just means i wont be buying any Nvidia products until they do. I fucking LOVE my 750 ti, I want to see more cards that arent giant heat engines.
 
pervert said:
To see intel's fabbing to be hugely superior to anyone else. nvidia (and lol amd) should rent some 22nm or 14nm (soon) time from intel. tsmc and globalfoundries need to step up their game
Click to expand...

Couldn't Nvidia and AMD just build their own facilities? Or is it too expensive?
 
DASHlT said:
Something was wrong with TSMC and 20nm. Can't remember where I read up on it though.

Which sucks because I was REALLY looking forward to 20nm high end Maxwell...
Click to expand...

We'll just have to "settle" for 28nm Maxwell, which (if we're honest), isn't the worst thing in the world. :)
 
melk said:
Who cares really? It seems the real performance comes from the underlying architecture changes in Maxwell, no?
Click to expand...
Yes, but there are some very real die size obstacles scaling that design up. GM107 is power efficient but large.
 
jbltecnicspro said:
We'll just have to "settle" for 28nm Maxwell, which (if we're honest), isn't the worst thing in the world. :)
Click to expand...

Well it all depends. Sure the 750ti is a great card. But the bigger dies is whole other ball game.

But the best way to test a new GPU is by releasing a lower end version. Worked great for AMD.

Also the rumor is right now is Nvidia wont release the big die Maxwell right away. It will be mid-ranged maxwell just like the 680 was.

Interesting times ahead for sure.
 
melk said:
Who cares really? It seems the real performance comes from the underlying architecture changes in Maxwell, no?

750 Ti has shown what it can do @ 28nm with a 128bit bus. I'll gladly be upgrading to whatever big Maxwell comes out, don't really care about it being on 28nm.
Click to expand...

Same process node means the same speed unless there are architectural improvements. Architectural improvements mean more lanes, more circuits. This drives up the cost considerably as the node failure rate for a given chip increases exponentially.

In other words, sure you can add more circuits, but it will cost you. So high end parts will be unaffordable to most, and bragging rights at best.
 
jbltecnicspro said:
Couldn't Nvidia and AMD just build their own facilities? Or is it too expensive?
Click to expand...

The fab machines themselves are expensive. Then comes retooling and/or buying new machines every time the process changes. Couple in that they need clean rooms - I believe they're something like 100,000 times more cleaner than the air we breathe. Not likely a cheap item, especially for just making GPU's.

AMD has a better shot at doing it. Since they also make processors, they have a larger market sector.

Tbh, I'm surprised Intel just doesn't buy Nvidia.
 
DigitalGriffin said:
Same process node means the same speed unless there are architectural improvements. Architectural improvements mean more lanes, more circuits. This drives up the cost considerably as the node failure rate for a given chip increases exponentially.

In other words, sure you can add more circuits, but it will cost you. So high end parts will be unaffordable to most, and bragging rights at best.
Click to expand...

What is node failure rate?

If Maxwell arch adds more lanes, more circuits etc... isn't a 28nm die size actually beneficial by having more room?
 
Spidey329 said:
The fab machines themselves are expensive. Then comes retooling and/or buying new machines every time the process changes. Couple in that they need clean rooms - I believe they're something like 100,000 times more cleaner than the air we breathe. Not likely a cheap item, especially for just making GPU's.

AMD has a better shot at doing it. Since they also make processors, they have a larger market sector.

Tbh, I'm surprised Intel just doesn't buy Nvidia.
Click to expand...
GlobalFoundries was AMD. AMD spun them off in 2009. If it was more economical for AMD to own their own fabs, I'm sure they still would. The problem is, fabs are extremely expensive, the R&D to keep them current is extremely expensive, and AMD alone doesn't have the capital to do this. Neither does NVIDIA, most likely.

I believe that there are no 20nm high-performance fabs right now. Both the TSMC and the GF foundries are doing 20nm low power, which is where most of the money is anyways with mobile products, and we won't see a high power node until 16nm/14nm with FINFET.
 
melk said:
What is node failure rate?

If Maxwell arch adds more lanes, more circuits etc... isn't a 28nm die size actually beneficial by having more room?
Click to expand...

The node failure rate is the percentage of chips and boards that are bad based on a given mfg process. The # of good chips on a given wafer starts to decrease exponentially as the die size increases linearly. For example, if you have a 5% increase in die size due to more circuits, the number of useable chips might decrease by 30% because it gets harder and harder to have a chip that is defect free with a larger and larger area.

(Think of it like a bag full of 1000 blue balls and 1 red ball (the defect)) Pull out 300 or so of those balls, and the odds are in your favor that they will all be blue. After that you are pushing your luck.

The formula for estimating success rate is well known. Mfgs will not switch to a new node until the # of useable chips/wafer is roughly as same as the old process.

But given two chips of the same process node size, the smaller chip will always be SIGNIFICANTLY cheaper.
 
DigitalGriffin said:
The node failure rate is the percentage of chips and boards that are bad based on a given mfg process. The # of good chips on a given wafer starts to decrease exponentially as the die size increases linearly. For example, if you have a 5% increase in die size due to more circuits, the number of useable chips might decrease by 30% because it gets harder and harder to have a chip that is defect free with a larger and larger area.

(Think of it like a bag full of 1000 blue balls and 1 red ball (the defect)) Pull out 300 or so of those balls, and the odds are in your favor that they will all be blue. After that you are pushing your luck.

The formula for estimating success rate is well known. Mfgs will not switch to a new node until the # of useable chips/wafer is roughly as same as the old process.

But given two chips of the same process node size, the smaller chip will always be SIGNIFICANTLY cheaper.
Click to expand...
A smaller die size also enables them to get more chips per wafer which reduces costs substantially. Wafers are a fixed size - if your die size decreases then the # of usable chips per wafer go up. Mfgs typically pay for X amount of wafers, not the individual chips, which is why this is significant.
 
DigitalGriffin said:
The node failure rate is the percentage of chips and boards that are bad based on a given mfg process. The # of good chips on a given wafer starts to decrease exponentially as the die size increases linearly. For example, if you have a 5% increase in die size due to more circuits, the number of useable chips might decrease by 30% because it gets harder and harder to have a chip that is defect free with a larger and larger area.
Click to expand...

All of that gets worked out once they get the libraries from the fab and they are designing the ASIC. They build in enough redundancy to get acceptable yields.
 
So is there any reliable evidence of when we will see the expanded Maxwell lineup? I don't really care if it is at 28nm or 20nm as long as there is a continuation of the kind of performance/watt we've seen with the 750/750 Ti.
 
The GK110 could benefit more ROPs and buswidth at 4K, for sure. And if they can ditch some of those "wasted" compute transistors that don't benefit gaming to squeeze more cores out of 28nm GM, I'm up for it.
 
MrCaffeineX said:
So is there any reliable evidence of when we will see the expanded Maxwell lineup? I don't really care if it is at 28nm or 20nm as long as there is a continuation of the kind of performance/watt we've seen with the 750/750 Ti.
Click to expand...
How long does it take a graphics card to reach market after being taped out?
4-6 months? There's some reliable evidence.
 
TaintedSquirrel said:
How long does it take a graphics card to reach market after being taped out?
4-6 months? There's some reliable evidence.
Click to expand...
I think the fastest it goes from tape out to market is 3-4 months. Assuming that they don't have to do any respins to fix issues (since 28nm is very mature at this point) I would say we should see NVIDIA's next iterations around September/October.
 
Rizen said:
A smaller die size also enables them to get more chips per wafer which reduces costs substantially. Wafers are a fixed size - if your die size decreases then the # of usable chips per wafer go up. Mfgs typically pay for X amount of wafers, not the individual chips, which is why this is significant.
Click to expand...

Not necessarily. Remember as the node gets smaller so do the circuit pathways. While the pathway shrinks, the wafer defect rate stays the same. This is particularly disadvantagous with posson distribution of defects which says wafer defects can cluster together. This leads to speed binning sometimes because a couple circuit paths dont make the grade because of defects in the path which slows or stops the circuit entirely.

2 or 3 missed placed atoms on 20 atom wide path isnt nearly as devasting as 2 or 3 misplaced atoms on a 10 atom wide path.


This is why yields dont improve with smaller process nodes despite an increase in the number of chips per waffer.
 
It's sad that only Intel seems to really invest in fab technology at a substantial rate. All the other companies scaled back their R&D on fabs during the downturn, except Intel. So here they are, doing 22nm for two years while nobody else can, and they have 14nm fabs online, but are delaying it because, heck, they don't need to push forward given the lack of competition. They are already a node ahead.
 
Sycraft said:
It's sad that only Intel seems to really invest in fab technology at a substantial rate. All the other companies scaled back their R&D on fabs during the downturn, except Intel. So here they are, doing 22nm for two years while nobody else can, and they have 14nm fabs online, but are delaying it because, heck, they don't need to push forward given the lack of competition. They are already a node ahead.
Click to expand...

Intel wouldn't delay 14nm. They need it for the tablet and soc market to compete with ARM
 
Sycraft said:
It's sad that only Intel seems to really invest in fab technology at a substantial rate. All the other companies scaled back their R&D on fabs during the downturn, except Intel. So here they are, doing 22nm for two years while nobody else can, and they have 14nm fabs online, but are delaying it because, heck, they don't need to push forward given the lack of competition. They are already a node ahead.
Click to expand...

20nm is humming along just fine.

But not for AMD and NVIDIA.
 
jwcalla said:
20nm is humming along just fine.

But not for AMD and NVIDIA.
Click to expand...
20nm low power processes are humming along fine. Graphics cards aren't made on the low power process though, they are made on the high performance process. There are no 20nm HP processes because they were not seeing any gains over 28nm, thus the delay.
 
You must log in or register to reply here.
Back
Top