NVIDIA Corporation CEO Says 16-Nanometer Pascal Yields Are "Good"

[Deleted User]

The user and all related content has been deleted.

Quizzical

Nvidia wouldn't announce anything if yields weren't good?  You mean like with Fermi, where they announced a bunch of products and even launched a bunch of cards with widespread availability, but never did offer commercial products with fully functional dies for either of the top two dies?  And for the third die down, the fully-functional chips were only used in laptops.  And they had to do a base layer respin for their entire lineup to fix yields, something that I'm not aware of any other discrete GPU ever getting.

And Nvidia doesn't want to go first on new process nodes anymore?  Well, they're likely the first one production high-power chips on TSMC 16 nm.  It's been many years since Nvidia had the first high-power chips on a node, as for many years, they'd wait to launch several months after AMD did on the same node.  They're also very early on FinFet Plus, which is a different node from Apple's early products, though they are related.

There's also the issue of Nvidia claiming about 35 GFLOPS/W in a Tesla P100 and about 50 GFLOPS/W on a GeForce GTX 1080.  They're on the same process node, and the former is both clocked much lower and has HBM2, both of which should improve efficiency.  I could believe good yields on the GP104 chip for this early in a process node lifetime, but the public information on GP100 sure makes it look like it's really, really broken.

Malabooga

Please dont link articles from these investor sites because theyre ridiculous.

Both NVidia and AMD "skipped" 20nm bbecause it was disaster for high performance chips. Both NVidia and AMD had designs for 20 nm (Maxwell and Fiji) and were pretty much screwed by TSMC. As Quizz said its pretty much first high performance chip on TSMC 16nm and not "same" process (same as Samsungs 14nm LPE and 14nm LPP for AMD)

Yields are "good" can mean anything. How "good"? No numbers? YMMV Not producing chips on 16nm would mean still producing chips on 28nm....which means still producing Maxwell and Kepler.

Doug_B

The fool publication are hardliner Nvidia  / Intel investers. They would call any and everythig nvidia does amazing ,as they continue to write off AMD in every chance they get. So read their article with a grain of salt.

[Deleted User]

The user and all related content has been deleted.

Ridelynn

This is only interesting because we weren't terribly concerned with GP104 yields, we are more concerned over GDDR5X production. 

Apart from that, sounds like more spin aimed more at Wall Street than anything actually grounded in anything resembling reality. And given the source, I'd say that's almost exactly what we are looking at.

SEANMCAD

Ridelynn said:

This is only interesting because we weren't terribly concerned with GP104 yields, we are more concerned over GDDR5X production. 

Apart from that, sounds like more spin aimed more at Wall Street than anything actually grounded in anything resembling reality. And given the source, I'd say that's almost exactly what we are looking at.

I often dream of a day when a poster says 'I thought X would happens and as it turns out it didnt, oh well I was wrong but moving on'

almost never happens

Ridelynn

Hey, if it happens that I'm wrong and there's wide availability, I may not be the first to post it, but I will happily post that I was wrong.

But we aren't there just yet. Don't count your chickens.

Hrimnir

Don't know if you know Ridelynn, but Micron CEO said that the yields for GDDR5X are excellent and that they are in full blown production right now.  So, I honestly don't think that's gonna be a hindrance.  Also, from what Nvidia's CEO said today the yields are good for GP104, so, unless that's a bull face lie, I don't think early july for widespread availability is out of the question.

Quizzical

Hrimnir said:

Don't know if you know Ridelynn, but Micron CEO said that the yields for GDDR5X are excellent and that they are in full blown production right now.  So, I honestly don't think that's gonna be a hindrance.  Also, from what Nvidia's CEO said today the yields are good for GP104, so, unless that's a bull face lie, I don't think early july for widespread availability is out of the question.

When Amazon or New Egg or whoever accepts a bunch of video cards into their warehouse, the silicon inside doesn't come from wafers that started the week before.  It could easily be six months before.  If yields are excellent today, then that means there should be widespread availability several months from now.  Not tomorrow, next week, or even next month, unless yields have been excellent for quite some time and mass production started a number of months ago.

Good yields are a relative thing.  Yields that would be considered fantastic for the first chips on a new node could get people fired for older chips on very mature nodes.  As both the chip designer and the fab have time to see what is wrong and tweak things, yields tend to go up with time.  It would be absolutely staggering if yields on GP104 aren't much, much worse than yields on GM204 (GeForce GTX 970/980)--in part because yields on the latter are probably really, really good by now.  It would also be shocking if yields on GP104 aren't much better a year from today than they are today.  The same is true of Polaris 10 and 11, Zen, and any other new or upcoming chips on new process nodes.  None of that contradicts the claim that GP104 yields are good for being so early in the life cycle of both the chip and the process node.

Furthermore, once yields are good enough to justify mass production, better or worse yields don't impact availability that much.  The difference between being able to sell 70% of your dies versus 95% isn't that big of a difference in how many cards come to retail--and might be no difference at all if the former leads you to buy more wafers in order to get the good dies you need.  It is, however, a huge deal for profitability.  Designers of CPUs and GPUs with huge gross profit margins on a per chip basis may be able to eat increased production costs from poor yields to some degree, but makers of commodity DRAM can't do so for long if they hope to stay in business.

No one doubts that there will be widespread availability of the GeForce GTX 1070 and GTX 1080 eventually, apart from perhaps the cranks who expect some cataclysmic end of civilization any day now.  It would be very surprising if you can't find a bunch of different GTX 1080s for $600 or below at the start of 2017.  But it's only a question of when, and widespread availability of video cards using a memory standard that only entered sampling a mere two months earlier would be unprecedented.

Hrimnir

@quiz there were reports from February that the early samples that micron had were already clocking excellent and that yields were good at that point.  I think you're stretching here.  Either way we're at the point that time will tell.  I promise I won't rub it in too hard If I was right.  And if I'm wrong and you're right i'll gladly eat crow.

https://www.micron.com/about/blogs/2016/february/gddr5x-has-arrived 

February 09, 2016

The team at our Graphics DRAM Design Center in Munich, Germany is doing a fantastic job, too. Not only do we have functional devices earlier than expected, these early components are performing at data rates of more than 13Gb/s! Memory components mature as they move through the development and manufacturing process, so to see first silicon performing at nearly full performance specs was a pleasant surprise—these early results are incredibly promising

Quizzical

Hrimnir said:

@quiz there were reports from February that the early samples that micron had were already clocking excellent and that yields were good at that point.  I think you're stretching here.  Either way we're at the point that time will tell.  I promise I won't rub it in too hard If I was right.  And if I'm wrong and you're right i'll gladly eat crow.

https://www.micron.com/about/blogs/2016/february/gddr5x-has-arrived 

February 09, 2016

The team at our Graphics DRAM Design Center in Munich, Germany is doing a fantastic job, too. Not only do we have functional devices earlier than expected, these early components are performing at data rates of more than 13Gb/s! Memory components mature as they move through the development and manufacturing process, so to see first silicon performing at nearly full performance specs was a pleasant surprise—these early results are incredibly promising

What you don't seem to realize is that the evidence you present not only doesn't support your conclusion, but points in the opposite direction.

Let's suppose that AMD or Nvidia decided to make a new discrete GPU chip and just got back the first wafers from it today.  By some miracle, on the first wafers they get back, not only is every single die fully functional, but they all clock above the target clock speeds.  When would they be able to do a hard launch with this new GPU, assuming they can get as many wafers as they want and there is an abundant supply of everything except the GPU dies?  The answer there is probably somewhere around October or November.  Not next week, or even two months later.

Hrimnir

Your argument was initially that GDDR5X production would hold back mass availability of the cards. I've given solid information to counter that argument and that both retailers and NVidia likely have excellent supply of gddr5x within the next 2-3 weeks to produce cards.

The second argument was that due to the new process, etc, yields for the GP104 wouldn't be great and it will be 4-6 months or more before the cards are readily available.  Nvidia's CEO came out and said that yields are good, which again, unless you're willing to state that as a full blown lie, then that defeats that argument.

So to get back to what you said, let's take your exact statement, they got "excellent wafers" today, then yes, October or November would be ideal.  However, we are talking about micron having those excellent wafers in early feb which is 3.5 months ago, so 1.5 months from now puts us at the start of july, even being generous and giving it an extra month puts it at 6 months from initial "good" wafers which is PRECISELY in line with your argument of 4-6 months.

As for NVidia, we have no idea when they started proper mass production of GP100/GP104. You ASSUME it was recently, however it could have been months ago.

At the end of the day, the only way either of us will know is in due time.  So, again, I'm prepared to eat crow if I'm wrong, are you?

SEANMCAD

ummm...yeah

Mayb 27th, $599, Microcenter, GDDR5X

Quizzical

Hrimnir said:

So to get back to what you said, let's take your exact statement, they got "excellent wafers" today, then yes, October or November would be ideal.  However, we are talking about micron having those excellent wafers in early feb which is 3.5 months ago, so 1.5 months from now puts us at the start of july, even being generous and giving it an extra month puts it at 6 months from initial "good" wafers which is PRECISELY in line with your argument of 4-6 months.

Better than you expect from first silicon is a long, long way away from being absolutely flawless.  If you get back perfect wafers, you get on the phone and order a big production run immediately.  If you merely get back wafers that are better than you expect from first silicon, maybe you get lucky and only have to do one respin, which means a delay of about two months.  That's not a "product is late" sort of delay; it's something that you budget for when creating a timetable, which is why Micron had originally planned to enter mass production of GDDR5X this summer.

Apparently things went well for Micron, so they were able to start mass production of GDDR5X in May rather than July or August or so.  But you don't expect mass availability of a product that entered production earlier that month, especially when wafers take several weeks to go through the fabs.  Well, maybe you do, but you shouldn't.

Quizzical

SEANMCAD said:

ummm...yeah

Mayb 27th, $599, Microcenter, GDDR5X

There's a big difference between:

a)  a sign on a shelf that says out of stock and we'll let you get on a waiting list, and
b)  enough stock that everyone who wants one for $599 can easily find one at that price or below.

The reason I said "mass availability" as meaning New Egg earlier was precisely because if something is out of stock, New Egg will very clearly mark it as out of stock--and they'll take the listing down out of searches if it's been out of stock entirely for two weeks.  A lot of sites will try to make it look like something is in stock to get you to order it.

SEANMCAD

Quizzical said:

SEANMCAD said:

ummm...yeah

Mayb 27th, $599, Microcenter, GDDR5X

There's a big difference between:

a)  a sign on a shelf that says out of stock and we'll let you get on a waiting list, and
b)  enough stock that everyone who wants one for $599 can easily find one at that price or below.

The reason I said "mass availability" as meaning New Egg earlier was precisely because if something is out of stock, New Egg will very clearly mark it as out of stock--and they'll take the listing down out of searches if it's been out of stock entirely for two weeks.  A lot of sites will try to make it look like something is in stock to get you to order it.

lol...

yeah like Ninvida has no experience with this.

How much you want to wager I will have a 1080 in my hands before the end of June?

Quizzical

Hrimnir said:

The second argument was that due to the new process, etc, yields for the GP104 wouldn't be great and it will be 4-6 months or more before the cards are readily available.  Nvidia's CEO came out and said that yields are good, which again, unless you're willing to state that as a full blown lie, then that defeats that argument.

As for NVidia, we have no idea when they started proper mass production of GP100/GP104. You ASSUME it was recently, however it could have been months ago.

I never argued that production of GP104 chips was certain to be a major source of delays.  I argued that we don't know--and furthermore, we still don't.  It's possible that the GTX 1070 will do a hard launch and the "founder's edition" cards will be irrelevant because everyone wants to save $70.  It's possible that it will be a soft launch with the $450 version the only thing you can find for a month.  It's possible that it will be a paper launch with virtually nothing available for a month.  We don't know.

You can make nearly the same statements about Polaris 10 and 11, too, with some slight tweaks for the prices.  Though with Polaris shown off publicly in January, AMD has had time to fix whatever was broken by now.  How soon Global Foundries would have fixed whatever is wrong with their process node is another issue entirely, and their history on this is substantially worse than TSMC.

I'm fairly certain that Nvidia didn't start mass production of GP104 before January, as if they had, they could have shown off real, working GPUs at CES instead of fake ones.  But starting mass production in February would be consistent with a hard launch in June.  Starting mass production a month ago would not.  We don't know when they started.

As for GP100, I'd be very surprised if Nvidia has started mass production even today.  From their claims of 35 GFLOPS/W on a Tesla P100 and 50 GFLOPS/W on a GTX 1080, it sure looks like GP100 is really, really broken.  Now, there is time to fix it, even if they have to do a base layer respin, as it would end up waiting on HBM2, anyway.  And I'm confident that they will mostly fix it.  But since they knew that it's horribly broken and they'd have time to fix it without really delaying launch much because you're waiting on memory, starting mass production of chips that you know are horribly broken would be a completely insane thing to do.

I don't know when they had their first silicon come back from GP100.  Maybe they'll have a fixed version ready to start mass production in a month or two, or maybe it will take longer.  But GP100 is still an early 2017 product, or at best, late 2016.  Even if they did start mass production on broken GP100 chips, you don't want them, as the GTX 1080 would be a far superior option and likely available sooner.

Quizzical

SEANMCAD said:

Quizzical said:

SEANMCAD said:

ummm...yeah

Mayb 27th, $599, Microcenter, GDDR5X

There's a big difference between:

a)  a sign on a shelf that says out of stock and we'll let you get on a waiting list, and
b)  enough stock that everyone who wants one for $599 can easily find one at that price or below.

The reason I said "mass availability" as meaning New Egg earlier was precisely because if something is out of stock, New Egg will very clearly mark it as out of stock--and they'll take the listing down out of searches if it's been out of stock entirely for two weeks.  A lot of sites will try to make it look like something is in stock to get you to order it.

lol...

yeah like Ninvida has no experience with this.

How much you want to wager I will have a 1080 in my hands before the end of June?

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

Look at the date there.  Cards didn't actually show up until May of that year.  Yep, Nvidia has experience with paper launches.  And that was in spite of AMD having widespread availability using the same memory and process in January of that year.  The GeForce GT 240 (Nvidia's first GDDR5 GPU) was about a four month delay between the first cards being available and widespread availability for anyone who foolishly wanted to buy one.

Quizzical

SEANMCAD said:

How much you want to wager I will have a 1080 in my hands before the end of June?

If Nvidia has sold N 1080s by the end of June, and they're universally out of stock or at higher prices because there are 5N people who would have bought one if they could for $599, then I'm going to declare that I was right even if you're one of the N people who managed to get one.

SEANMCAD

Quizzical said:

SEANMCAD said:

Quizzical said:

SEANMCAD said:

ummm...yeah

Mayb 27th, $599, Microcenter, GDDR5X

There's a big difference between:

a)  a sign on a shelf that says out of stock and we'll let you get on a waiting list, and
b)  enough stock that everyone who wants one for $599 can easily find one at that price or below.

The reason I said "mass availability" as meaning New Egg earlier was precisely because if something is out of stock, New Egg will very clearly mark it as out of stock--and they'll take the listing down out of searches if it's been out of stock entirely for two weeks.  A lot of sites will try to make it look like something is in stock to get you to order it.

lol...

yeah like Ninvida has no experience with this.

How much you want to wager I will have a 1080 in my hands before the end of June?

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

Look at the date there.  Cards didn't actually show up until May of that year.  Yep, Nvidia has experience with paper launches.  And that was in spite of AMD having widespread availability using the same memory and process in January of that year.  The GeForce GT 240 (Nvidia's first GDDR5 GPU) was about a four month delay between the first cards being available and widespread availability for anyone who foolishly wanted to buy one.

how much do you want to wager. for giggles fake money of course. or if you like we can meet in person at a Microcenter, lets say June 15

SEANMCAD

Torval said:

Sean, just post a pic of it next to your DK2. That will be enough.

LOL...nicely done I like that. more exposure to that the better for me.

I did get an email from Microcenter saying GTX 1080 in stores May 27. I suspect they will be common place in most major cities by around mid June

Hrimnir

Quizzical said:

Hrimnir said:

So to get back to what you said, let's take your exact statement, they got "excellent wafers" today, then yes, October or November would be ideal.  However, we are talking about micron having those excellent wafers in early feb which is 3.5 months ago, so 1.5 months from now puts us at the start of july, even being generous and giving it an extra month puts it at 6 months from initial "good" wafers which is PRECISELY in line with your argument of 4-6 months.

Better than you expect from first silicon is a long, long way away from being absolutely flawless.  If you get back perfect wafers, you get on the phone and order a big production run immediately.  If you merely get back wafers that are better than you expect from first silicon, maybe you get lucky and only have to do one respin, which means a delay of about two months.  That's not a "product is late" sort of delay; it's something that you budget for when creating a timetable, which is why Micron had originally planned to enter mass production of GDDR5X this summer.

Apparently things went well for Micron, so they were able to start mass production of GDDR5X in May rather than July or August or so.  But you don't expect mass availability of a product that entered production earlier that month, especially when wafers take several weeks to go through the fabs.  Well, maybe you do, but you shouldn't.

If we were talking about CPU or GPU you would be correct, but we're not, memory is relatively uncomplicated compared to making dies for CPUs and GPUs.  Memory is cheap for a reason.

Malabooga

Quizzical said:

Hrimnir said:

The second argument was that due to the new process, etc, yields for the GP104 wouldn't be great and it will be 4-6 months or more before the cards are readily available.  Nvidia's CEO came out and said that yields are good, which again, unless you're willing to state that as a full blown lie, then that defeats that argument.

As for NVidia, we have no idea when they started proper mass production of GP100/GP104. You ASSUME it was recently, however it could have been months ago.

I never argued that production of GP104 chips was certain to be a major source of delays.  I argued that we don't know--and furthermore, we still don't.  It's possible that the GTX 1070 will do a hard launch and the "founder's edition" cards will be irrelevant because everyone wants to save $70.  It's possible that it will be a soft launch with the $450 version the only thing you can find for a month.  It's possible that it will be a paper launch with virtually nothing available for a month.  We don't know.

You can make nearly the same statements about Polaris 10 and 11, too, with some slight tweaks for the prices.  Though with Polaris shown off publicly in January, AMD has had time to fix whatever was broken by now.  How soon Global Foundries would have fixed whatever is wrong with their process node is another issue entirely, and their history on this is substantially worse than TSMC.

I'm fairly certain that Nvidia didn't start mass production of GP104 before January, as if they had, they could have shown off real, working GPUs at CES instead of fake ones.  But starting mass production in February would be consistent with a hard launch in June.  Starting mass production a month ago would not.  We don't know when they started.

As for GP100, I'd be very surprised if Nvidia has started mass production even today.  From their claims of 35 GFLOPS/W on a Tesla P100 and 50 GFLOPS/W on a GTX 1080, it sure looks like GP100 is really, really broken.  Now, there is time to fix it, even if they have to do a base layer respin, as it would end up waiting on HBM2, anyway.  And I'm confident that they will mostly fix it.  But since they knew that it's horribly broken and they'd have time to fix it without really delaying launch much because you're waiting on memory, starting mass production of chips that you know are horribly broken would be a completely insane thing to do.

I don't know when they had their first silicon come back from GP100.  Maybe they'll have a fixed version ready to start mass production in a month or two, or maybe it will take longer.  But GP100 is still an early 2017 product, or at best, late 2016.  Even if they did start mass production on broken GP100 chips, you don't want them, as the GTX 1080 would be a far superior option and likely available sooner.

GP100 failed before it begun

https://techaltar.com/google-made-asic-processor-ai/

SEANMCAD

Malabooga said:

GP100 failed before it begun

https://techaltar.com/google-made-asic-processor-ai/

now THAT is troll bait

Ridelynn

Uh oh

http://www.eteknix.com/nvidia-gtx-1080-pre-orders-sell-out-in-minutes/

I'm sure that's great that it "sells out more often than not" - but it also is not a good data point to have coming up on an impending wide-availability public release. 

It's not May 27, so not saying anything apart from the fact that this isn't a good sign for wide availability. They still have a week to make sure there is sufficient stock available.