TSMC's R&D boss addresses 40-nm yields, high-k, litho

[Nedjo]

well qcmadness it looks like you've slapped me twice in a row! hats down to you

[saaya]

wow, so ati HAS to use GF? that might actually cause them trouble if theres a problem with GF?

[Mechanical Man]

Actually, when you read it carefully. You see AMD has to ramp their GPU production in 5 years. Eg. Now they are starting with Lliano (integrated CPU+GPU) and during rest five years they do need to ramp up the production when viable. So if GF cant supply good process for AMD GPUs, AMD does not need to move their GPUs to GF. Also, there is no mention on that exerpt about what end percentage should even be. There is space for discrete graphics staying on TSMC. Remember, most high volume products are integrated GPUs..

[Olivon]

Digitimes - GPU shortage unlikely to be solved before May

The current shortage of graphics chips, which is being caused largely by low yields of the 40nm process at Taiwan Semiconductor Manufacturing Company (TSMC), is unlikely to be completely solved before May, according to sources from graphics card makers.

Although TSMC recently said the defect density of its 40nm technology has already dropped from 0.3-0.4 per square inch to 0.1-0.3, the sources pointed out that the improvement in overall yield still needs more time before catching up with market demand.

To maintain gross margins from the increased cost created by the GPU shortages and rising DRAM quotes, graphics card makers including Asustek Computer, Micro-Star International (MSI), Gigabyte Technology and Tul are planning to increase prices of their products by 5-10%, according to market sources.

[Sn0wm@n]

TSMC has been producting .40nm chips for ATi months before consumers were able to purchase them... and much of those 1 million chips were manufactured before December while yields were very low. This is why I've always maintained the opinion that ATi jumped the gun just to get the crown for a while but forgot to remember that without profits, they wont be able to pay off all the R&D they're spending on very fast product cycles.

this makes sense....


so ati producing chips and getting them to market and having a good hold on the market for now is bad to recoup the R&D cost ???? sure......

[Nedjo]

If TSMC is not able to fulfill orders for both ATi and nVidia, they will have to produce chips for the company they have a more solid contract with. My bet would be on nVidia since they have far more market share and are a much larger source of income for TSMC.

Now let's face the facts:
a)


those two guys looks to be happy working together!

b)
AMD has no issues with supply of 40nm DX11 chips!

c)
There's one another company that can't get enough 40nm DX11 chips

d)
TSMC boss firmly stated in couple of occasions that they've fixed problems with 40nm, and yields are fine, with continuous ramping of the output wafers!

b+c+d = stories about issues with TSMC's 40nm tech are coming from the side that has broken design for their massive 40nm chip!

So what's the measure of TSMC's success??
a) they can successfully build 2 billion transistors chip on 40nm
b) they can't successfully build 3 billion chip on 40nm

SO POINT:
what do you think that TSMC will chose?? To have their equipment producing what they CAN produce, or to have them idle 'cos they can't produce 3 Billion chip with sufficient yield in sufficient quantities??

TSMC has been producting .40nm chips for ATi months before consumers were able to purchase them... and much of those 1 million chips were manufactured before December while yields were very low. This is why I've always maintained the opinion that ATi jumped the gun just to get the crown for a while but forgot to remember that without profits, they wont be able to pay off all the R&D they're spending on very fast product cycles.

yeah those PhD and MBA 'tards in ATI don't get the simple fact that they can get paycheck only if they make profit! Please be humane and send them a notice, 'cos you're obviously aware of that simple fact!!

nVidia has almost 3x the market share of ATi. A 'low volume' card for nVidia is probably quite similar in units produced as a 'high volume' card from ATi.

And you have some source to back up this brilliant statement??

Plus I have a feeling TSMC makes more profit from larger accounts like nVidia.

well if you have feeling who're wee to confront you

[saaya]

Actually, when you read it carefully. You see AMD has to ramp their GPU production in 5 years. Eg. Now they are starting with Lliano (integrated CPU+GPU) and during rest five years they do need to ramp up the production when viable. So if GF cant supply good process for AMD GPUs, AMD does not need to move their GPUs to GF. Also, there is no mention on that exerpt about what end percentage should even be. There is space for discrete graphics staying on TSMC. Remember, most high volume products are integrated GPUs..

k... so they basically HAVE to keep a foot in the door at GF and give them a try, so to speak... hoping that as soon as GF does a good job or TSMC stumbles they are ready to go full throttle @ GF... makes sense... but wouldnt they do this either way? its weird that they mention this in a contract, isnt it?

its kinda like signing a contract that tells you that you HAVE to look for other job opportunities that might be more fun or pay you a better salary... well duh! who doesnt do that?

Digitimes - GPU shortage unlikely to be solved before May

thats interesting!

VERY interesting!

0.1-0.3 errors per square inch...
1 square inch = 25.4x25.4mm
fermi =~ 24x24mm

so on average there is 1 error per 3-10 fermi chips?
that doesnt make any sense... at all...
if this would be true then fermi should have yields of 70-90% fully functional 480 chips and 5-25% salvageable broken chips aka 470s, and rv870 should have yields of 80-95% fully functional and 5-15% salvageable parts aka 5850 or 5830

[wuttz]

amd/ati's agreements with tsmc&glofo is meant to address the "problem" of spintel escalating a node-process war. if you ask me, it is a brilliant short-term solution to keep up with tick-tock, freeing up amd/ati from manufacturing headaches and allowing them to fully concentrate on gpu/cpu/cgpu design. meanwhile glofo will be tip of the spear fighting the manuf war vs. spintel, with deep financial pockets courtesy of abudhabi.

im waiting to see how chipzilla counters this strategy. or will it be more of the same old trick bribing OEM's.

[Nedjo]

0.1-0.3 errors per square inch...
1 square inch = 25.4x25.4mm
fermi =~ 24x24mm

so on average there is 1 error per 3-10 fermi chips?
that doesnt make any sense... at all...
if this would be true then fermi should have yields of 70-90% fully functional 480 chips and 5-25% salvageable broken chips aka 470s, and rv870 should have yields of 80-95% fully functional and 5-15% salvageable parts aka 5850 or 5830

defects are not only thing that have influence on yield...

[Chumbucket843]

b)
AMD has no issues with supply of 40nm DX11 chips!

false.

TSMC boss firmly stated in couple of occasions that they've fixed problems with 40nm, and yields are fine, with continuous ramping of the output wafers!

b+c+d = stories about issues with TSMC's 40nm tech are coming from the side that has broken design for their massive 40nm chip!

So what's the measure of TSMC's success??
a) they can successfully build 2 billion transistors chip on 40nm
b) they can't successfully build 3 billion chip on 40nm

SO POINT:
what do you think that TSMC will chose?? To have their equipment producing what they CAN produce, or to have them idle 'cos they can't produce 3 Billion chip with sufficient yield in sufficient quantities??

you cant have it both ways. ATi pays tsmc less money per chip and sells their products at a lower price point. it should be obvious that tsmc is making more money off of nvidia. hell, they probably payed off the majority of the cost of 40nm node.

[saaya]

defects are not only thing that have influence on yield...

so the real problem isnt defect density but what? transistor variance?

[saaya]

you cant have it both ways. ATi pays tsmc less money per chip and sells their products at a lower price point. it should be obvious that tsmc is making more money off of nvidia. hell, they probably payed off the majority of the cost of 40nm node.

neither ati nor nvidia pay tsmc per chip, they pay them per wafer afaik, thats what several articles mentioned at least... and even IF ati would pay them per chip, their chips are smaller and use up less wafers and capacity... so with the same amount of wafers tsmc would make roughly the same amount of money with either one...

its ironic that you said "you cant have it both ways" because you just claimed that tsmc prefers nvidia because they pay them more money per wafer/chip and thats why nvidia gets lower wafer/chip prices... lol wutt?

[Nedjo]

so the real problem isnt defect density but what? transistor variance?

well it looks so...

[Chumbucket843]

they do pay per wafer but since nvidia has bigger die and lower yields it cost more per chip which is what matters to consumers and nvidia's profit margin. i do remember seeing some research that some company did. gt200 was costed $50 per chip and rv770 was $30 per chip. nvidia probably does get a deal but its not like it makes up for the cost.

[generics_user]

false.

when you do such claims back them up and please step back from putting out one wrong statement after another

you can get every single card of the 5kseries right now both in europe and in the US:

all HD5K series are in stock in the EU

Newegg - where is the shortage???

NCIX - no shortage at all

[zalbard]

7000 chips good availability? dont think so.

They won't. They have other much higher volume products on 40nm to worry about.

Yep, it probably won't even be 7000 chips a month (if we can keep your 10% yield figure which is too low). This is not a lot by all means, because shipping millions of these chips (like ATI does) would take years.

Going forward, we plan to introduce 22 nanometer node about two years after we introduce 28 nanometer, so the first introduction we'd like to be in the Q3 of 2012, and this if for the high performance version. And followed by the low power version about the end of Q1 2013.

Lovely, probably a year later than Intel...

[saaya]

well it looks so...

i dont know... there is no source mentioning this or even hinting at it... everybody talks about defects... gate variation affects transistor performance and leakage, but it doesnt break chips...

they do pay per wafer but since nvidia has bigger die and lower yields it cost more per chip which is what matters to consumers and nvidia's profit margin. i do remember seeing some research that some company did. gt200 was costed $50 per chip and rv770 was $30 per chip. nvidia probably does get a deal but its not like it makes up for the cost.

i find that hard to believe... if gt200 is only 50$ how come the cheapest card based on gt200 comes in at above 200$?
even now that 55nm is mature and the 260 is under price pressure it doesnt sell for less than 225$, which makes me guess that the chip cost is above 50$... probably close to 75$...

[Nedjo]

i dont know... there is no source mentioning this or even hinting at it... everybody talks about defects... gate variation affects transistor performance and leakage, but it doesnt break chips...

well transistors capable of 50% of targeted speed might as well be defected from usability, or yield standpoint.

[Chumbucket843]

i find that hard to believe... if gt200 is only 50$ how come the cheapest card based on gt200 comes in at above 200$?
even now that 55nm is mature and the 260 is under price pressure it doesnt sell for less than 225$, which makes me guess that the chip cost is above 50$... probably close to 75$...

the other parts were expensive to. nvidia caught a break by using gddr3 which was fairly cheap. there are a lot of other expenses besides manufacturing. testing and packaging are very costly. silicon is actually cheap. if tsmc can pay off the initial cost of the fab then their margins are ridiculous. high end is not a lucrative market (arguably hardware altogether) but its necessary for reputation.

when you do such claims back them up and please step back from putting out one wrong statement after another

you can get every single card of the 5kseries right now both in europe and in the US:

all HD5K series are in stock in the EU

Newegg - where is the shortage???

NCIX - no shortage at all

hemlock is very hard to find.
http://www.newegg.com/Product/Produc...n=5970&x=0&y=0
not to mention its $700.
http://www.google.com/products?hl=en...e2=&lnk=prsugg
yeah, me and my wrong statements. i wouldnt mind have 2 of these in a crunching rig.

[saaya]

well transistors capable of 50% of targeted speed might as well be defected from usability, or yield standpoint.

how come almost all rv870s can clock past 1ghz then?
even the lowest rv870 part aka 5830 comes clocked higher than 800mhz...

if this would be such a big problem dont you think they would disable more units for a 470 than just 2? a 460 or 450 with more disabled blocks or lower clocks would still make a looot more sense than throwing away those gpus that have SOME blocks that cant clock that high...

the other parts were expensive to. nvidia caught a break by using gddr3 which was fairly cheap. there are a lot of other expenses besides manufacturing. testing and packaging are very costly.

yeah but here its gddr5 vs gddr5 and i dont think testing gf100 will cost more than testing rv870, right? packaging maybe... but i dont think thats enough to make a difference in retail prices... maybe 5$ more for a more complex package?

http://www.zipzoomfly.com/jsp/Produc...odlist=froogle
wow, 599 for a 5970, thats a good offer for anybody interested in that much performance...
thats 100$ cheaper than newegg...

[Nedjo]

how come almost all rv870s can clock past 1ghz then?
even the lowest rv870 part aka 5830 comes clocked higher than 800mhz...

if this would be such a big problem dont you think they would disable more units for a 470 than just 2? a 460 or 450 with more disabled blocks or lower clocks would still make a looot more sense than throwing away those gpus that have SOME blocks that cant clock that high...

Anand's article about Evergreen development is quite reviling on this subject. Basically ATI has calculated all of those risks, and made TSMC 40nm-friendly design (thanks to RV740 for that).
NVIDIA didn't do that breaking to many rules and pushing TSMC's manufacturing capabilities to the limit.

yeah but here its gddr5 vs gddr5 and i don't think testing gf100 will cost more than testing rv870, right? packaging maybe... but i don't think thats enough to make a difference in retail prices... maybe 5$ more for a more complex package?

it's hard to be precise on that matter w/o insight in all the factors that constitute final expense.

http://www.zipzoomfly.com/jsp/Produc...odlist=froogle
wow, 599 for a 5970, thats a good offer for anybody interested in that much performance...
thats 100$ cheaper than newegg...

good price indeed.

[saaya]

only 75$ more than a gtx295 and only 200$ more than a 5870 (399$)... definitely very nice for whoever needs that much power

i read the anandtech article... but what anandtech didnt mention is that nvidia DID tape out other 40nm chips before gt300...
most of them were cancelled because they didnt work at all i think, but all the 10.1 parts were taped out before gf100... so nvidia did have SOME idea of 40nm before taping out gf100...
and especially considering the obvious clockspeed issues nvidia ran into with their 10.1 40nm parts, id be very surprised if they didnt take transistor variations into account when designing fermi...

[Behemot]

how come almost all rv870s can clock past 1ghz then?
even the lowest rv870 part aka 5830 comes clocked higher than 800mhz...

It makes sense I think. If you look at HD 5830, it has TDP almost as high as HD 5870. This implicates the chips needs higher voltage even they are smaller (the parts which are cutted of are not supplied with power). So probably these are the chips with not so good transistors?

As for the price of GT200 based cards - look at manufacturing track of every goods you can think about, not only electronics. You'll discover the manufacturer of first raw material has only few % of the final price. Rest is how the thing travels and everybody who it pases takes his 10% part...

E.g. tea or coffee is good example: the farmer gets few bucks, but it could cost up to hundreds dollars on the stock market on the other hamisphere. So basically if the GT200 costs 50 USD, the 200 bucks final price is actually very low; that's the reason many NVIDIA AIB's were so suffering last year: not only low prices, but low sales too.

ADD// The VGA shortage will be really bad, at least it looks like it really will be. If you look at the market, it looks that GT200, G92 and RV7x0 just are not manufactured anymore (GT200 for sure, but others too I think). Everything is 40nm now and even if yealds were 100%, 80 000 wafers/quarter is not enough for whole RV8x0, GT(S)2x0 and Fermi at this demand.

[BenchZowner]

I haven't read the thread, I'm kinda tired, bored and still gotta do lots of things
So please excuse me if somebody else came up with the same thought here before I ( just ) did

TSMC's boss saying "We're doing great and we fixed this blah blah" without any real proof, etc isn't as valuable as Jen-Hsun H. now saying "We have a great product and everything is going fine" ?

[Nedjo]

I haven't read the thread, I'm kinda tired, bored and still gotta do lots of things
So please excuse me if somebody else came up with the same thought here before I ( just ) did

TSMC's boss saying "We're doing great and we fixed this blah blah" without any real proof, etc isn't as valuable as Jen-Hsun H. now saying "We have a great product and everything is going fine" ?

ofcourse it's not a same!

on TSMC's defense we have bunch of 40nm Evergreen based cards from the second wave (MSI, GA, ASUS, Sapphire, PowerColor...) - check out TechPowerup... so TSMC is printing out Evergreens like there's no tomorrow... also it looks like there's plethora of NV DX10.1 chips out there... but there's no Fermi yet!