AMD announces open source ray tracing at GDC *Interview update*

[JR88]

Just now, dalekphalm said:

Evil?

 

Hookay.

 

You're seriously going to say that, and keep a straight face, and use Intel and NVIDIA in the same sentence?

 

Frankly, neither Intel, nor NVIDIA, nor AMD are evil. They're companies, who's responsibility is to the shareholders, and is to make profit. They aren't altruistic, nor are they straight up evil.

 

All three have done anti-competitive practices in the past. So if AMD is evil, so is Intel and NVIDIA (which could well be correct), but it's definitely not an "either-or" type situation.

Actually You're Right they all have done some Anti-competitive Practices before. It does just Look bad in General...I guess their all Evil idk lol

[leadeater]

3 hours ago, Razor01 said:

 

It has to be that high since there is no real cut down parts yet for Volta, We are talking about one memory bank difference between V100 parts so far.

 

Every nV chip, only has one cut down part, so @ 50% yield, they will not make any money on the entire wafer.

The CEO of Nvidia stood on stage and said they get one usable die from a wafer at the launch of Volta.

[Dylanc1500]

17 minutes ago, leadeater said:

The CEO of Nvidia stood on stage and said they get one usable die from a wafer at the launch of Volta.

Which is why he also stated that it costs them roughly $1000 to make an entire card, if memory serves? Its been awhile since I went through it all..

[dalekphalm]

20 minutes ago, leadeater said:

The CEO of Nvidia stood on stage and said they get one usable die from a wafer at the launch of Volta.

Do we know how big of a wafer they use? The 300mm (diameter) monsters?

[leadeater]

4 minutes ago, Dylanc1500 said:

Which is why he also stated that it costs them roughly $1000 to make an entire card, if memory serves? Its been awhile since I went through it all..

Don't remember but it was something crazy expensive due to the super low yield. 

[Dylanc1500]

2 minutes ago, dalekphalm said:

Do we know how big of a wafer they use? The 300mm (diameter) monsters?

12" wafer.

[dalekphalm]

24 minutes ago, Dylanc1500 said:

12" wafer.

Interesting.

 

With a 150mm radius (using a 300mm diameter standard wafer), you have a 70,685mm² area.

 

The die of a V100 is 815mm².

 

Using perfect efficiency of the surface area (impossible), you could fit 86 dies onto the wafer. Let's assume, because circles and squares (you know, Kindergarten stuff), that you can squeeze far less in. This side does a calculation for you:

http://www.silicon-edge.co.uk/j/index.php/resources/die-per-wafer

300mm wafer (all other settings left as default - I'm not a Wafer engineer so I wouldn't know what values to set anyway)

Assuming the die is a perfect square: 28.54mm x 28.54mm = ~815mm²

 

We get approximately 64 dies per wafer.

 

If - at launch - they were only getting ONE usable V100 per wafer, that's a 1.56% yield. That's... holy crap. That's bad.

 

Let's assume they've improved the yield, but even if they quadrupled the yield, we're still under 10% yield.

 

To get up to 60% yield, they would have to improve yields by 38 times.

[Razor01]

1 hour ago, leadeater said:

The CEO of Nvidia stood on stage and said they get one usable die from a wafer at the launch of Volta.

 

No he didn't say that, he stated fully functionally, two different things, and that was at launch. 

 

1 hour ago, leadeater said:

Don't remember but it was something crazy expensive due to the super low yield. 

https://www.theinquirer.net/inquirer/news/2256404/nvidia-claims-that-fabbing-big-chips-is-not-difficult

Quote

 

With Nvidia's Volta GPU architecture stacking DRAM on the same silicon substrate as the GPU, it will require the firm to increase the size of its chips. According to Sumit Gupta, GM of Nvidia's Tesla GPU Accelerated business unit, large chips are not hard to manufacture but rather tend to encounter difficulties passing verification.

 

 

And yeah the 1k for the entire card was stated at a financial call, that will give them if they are buying per wafer instead of good chip, 60% yields (@ launch)  Each wafer would be around 7k-8k.

 

But @ launch when selling at 18k per chip/card they can do what ever they want.

 

 

Also if you want to ask the guy I've talked to he is on B3D silent_guy is his username.  PM him, he works at TSMC.

 

I would link the thread at beyond 3d if the rules allow it, but already discussed it in depth there, its under the Volta Speculation thread.

[leadeater]

13 hours ago, Razor01 said:

No he didn't say that, he stated fully functionally, two different things, and that was at launch. 

At the time they only had one SKU, the V100. The statement wasn't quite what I said but it's very close to it, they are not or should I say at the time were not achieving 60% yields for V100 dies.

 

Quote

To make one chip work per 12 inch wafer I would characterize it as unlikely and so the fact that this is manufacturable is great it's just an incredible feat.

 

When your making statements that just getting one working is so hard then no matter how amazing TSMC is 60% simply is not realistic, that's potentially 38 V100 dies. That is not consistent at all with what is being said.

 

Your link is only a fluff piece before anything official about Volta was released, heck it's even 5 years old how is this even relevant to the discussion at all. That was a lead up piece for the GM200 601mm2 die which was very big for the time, which was also a testing ground for large die fabrication to validate if Volta V100 is even possible.

 

13 hours ago, Razor01 said:

And yeah the 1k for the entire card was stated at a financial call, that will give them if they are buying per wafer instead of good chip, 60% yields (@ launch)  Each wafer would be around 7k-8k.

Bill of materials cost isn't MSRP.

 

13 hours ago, Razor01 said:

Also if you want to ask the guy I've talked to he is on B3D silent_guy is his username.  PM him, he works at TSMC.

No thanks I'll take the CEO of Nvidia's statements over some guy on the internet.

[Razor01]

4 hours ago, leadeater said:

At the time they only had one SKU, the V100. The statement wasn't quite what I said but it's very close to it, they are not or should I say at the time were not achieving 60% yields for V100 dies.

 

 

When your making statements that just getting one working is so hard then no matter how amazing TSMC is 60% simply is not realistic, that's potentially 38 V100 dies. That is not consistent at all with what is being said.

 

Your link is only a fluff piece before anything official about Volta was released, heck it's even 5 years old how is this even relevant to the discussion at all. That was a lead up piece for the GM200 601mm2 die which was very big for the time, which was also a testing ground for large die fabrication to validate if Volta V100 is even possible.

 

Bill of materials cost isn't MSRP.

 

No thanks I'll take the CEO of Nvidia's statements over some guy on the internet.

 

 

Dude does that make financial sense?  I mean come on man, each wafer is 7 k

 

One good chip?  They are selling a V100 card for 3k

 

Come on.

 

They have at least 60% yields.  Its that simple.

 

If you want to make shit up, easy to do, but not so easy to do when you have the cost of the wafer!

 

The rest of what you stated doesn't fit with that AT ALL.

 

He was talking about a full working chip all parts fully functional. Yields aren't looked at that way, they need to have a certain amount of yields on all functioning chips period.  The lowest common denominator being the most prevalent because those are the volume parts.  With Volta, ya got pretty much 2 variants right now, and both have different parts cut.  That tells us if they are able to sell Volta at 3k a pop, they are getting more than one functional chip per die.  At least 2 if they are getting 0% profits!  Which oh wait right there is not what Jensen stated right if you think that is what he stated.

 

If nV cares about their margins, which we all know they do, they won't release such a card, ever!

 

Lets throw basic logic out the window when talking about money and see if it all adds up is that what we are getting at here?

 

Do you know why nV's margins when up ~3% last quarter even though they lost  marketshare?

 

It was because of Volta, you can expect they have much higher than 100% margins on this cards, because Volta is not a high volume piece, that 3% increase was caused by a huge increase in per product margins!  This is couple quarters after Volta was first shipped out to consumers, so what ever happened in the middle there, is what is causing the increase in margins.

[laminutederire]

11 minutes ago, Razor01 said:

 

 

Dude does that make financial sense?  I mean come on man, each wafer is 7 k

 

Once good chip?  They are selling a V100 card for 3k

 

Come on.

 

They have at least 60% yields.  Its that simple.

 

If you want to make shit up, easy to do, but not so easy to do when you have the cost of the wafer!

 

The rest of what you stated doesn't fit with that AT ALL.

 

He was talking about a full working chip all parts fully functional. Yields aren't looked at that way, they need to have a certain amount of yields on all functioning chips period.  The lowest common denominator being the most prevalent because those are the volume parts.  With Volta, ya got pretty much 2 variants right now, and both have different parts cut.  That tells us if they are able to sell Volta at 3k a pop, they are getting more than one functional chip per die.  At least 2 if they are getting 0% profits!  Which oh wait right there if not what Jensen stated right if you think that is what he stated.

 

If nV cares about their margins, which we all know they do, they won't release such a card, ever!

 

Lets throw basic logic out the window when talking about money and see if it all adds up is that what we are getting at here?

His point is simple: you can have good yields on your product stack, that doesn't mean you have 60 percent yields for the top end products. It's probable that they indeed are incapable of producing more than 10% of V100 dies. Yes they use damage dies to produce lower end products to recycle that out and but their losses on that. Fact is they probably make the least amount of margin on the biggest dies. It like Intel has serious trouble because the yields aren't good enough to have enough top end Xeons? Which is why they are so expensive. AMDs approach can help produce more functioning chips regardless of the size of the said chips. They have less to recycle and less absolutely useless dies thanks to that, which is why IF and it's successors do matter. If you want to produce your best products, it helps if you don't have to sell cut down versions as much. That's the point here.

[Razor01]

2 minutes ago, laminutederire said:

His point is simple: you can have good yields on your product stack, that doesn't mean you have 60 percent yields for the top end products. It's probable that they indeed are incapable of producing more than 10% of V100 dies. Yes they use damage dies to produce lower end products to recycle that out and but their losses on that. Fact is they probably make the least amount of margin on the biggest dies. It like Intel has serious trouble because the yields aren't good enough to have enough top end Xeons? Which is why they are so expensive. AMDs approach can help produce more functioning chips regardless of the size of the said chips. They have less to recycle and less absolutely useless dies thanks to that, which is why IF and it's successors do matter. If you want to produce your best products, it helps if you don't have to sell cut down versions as much. That's the point here.

 

 

Oh my god there are only 2 variants of Volta right now!  Both have some SM's cut and one has one memory bank cut too.

 

Not much difference at all.

 

its impossible to sell a Volta card at 3k, if they only get one functional chip out of a wafer!

 

Even with 100% yields, we are looking at a card that will cost them 700 bucks or so to make, there is just no way to get lower than that with 16gb of HBM2, and this die size.

[laminutederire]

3 minutes ago, Razor01 said:

 

 

Oh my god there are only 2 variants of Volta right now!  Both have some SM's cut and one has one memory bank cut too.

 

Not much difference at all.

 

its impossible to sell a Volta card at 3k, if they only get one functional chip out of a wafer!

The titan V is the cheapest one of the two... and it sells at 3000€. Imagine the price of the best variant. It costs more than a 8k$ wafer alone. HPE can sell you one at 10 to 20k$ (Product listing ). Nvidia sells four of them for more than 50k$. Those cards are profitable. For all we know, they sell the titan V because either their yields have came from one to maybe 2 or 3 cards per wafers, or because they produced too much Tesla cards they can't sell so since they're already produced might as well sell them.

 

[Razor01]

6 minutes ago, laminutederire said:

The titan V is the cheapest one of the two... and it sells at 3000€. Imagine the price of the best variant. It costs more than a 8k$ wafer alone. HPE can sell you one at 10 to 20k$ (Product listing ). Nvidia sells four of them for more than 50k$. Those cards are profitable. For all we know, they sell the titan V because either their yields have came from one to maybe 2 or 3 cards per wafers, or because they produced too much Tesla cards they can't sell so since they're already produced might as well sell them.

 

 

 

They will not sell a 3k card if they are making 0 bucks out of them, period.  They would not even release the card.  At 2 or 3 good GPU's per wafer, they will make no money.

 

The memory alone is 400 to 500 bucks man for them to buy.  Each stack of HBM 2, 4gb, is anywhere from 120 to 180 bucks.

[laminutederire]

2 minutes ago, Razor01 said:

 

 

They will not sell a 3k card if they are making 0 bucks out of them, period.  They would not even release the card.  At 2 or 3 good GPU's per wafer, they will make no money.

 

The memory alone is 400 to 500 bucks man for them to buy.  Each stack of HBM 2, 4gb, is anywhere from 120 to 180 bucks.

If you want. You're not listening to people so I'm not going to reiterate after this time:

If you have two good gpus, with one you sell at 10 to 20k$, then the other at 3k$, you still make profit and most importantly: you sell something that you wouldn't sell otherwise because they have already sold most of the Tesla cards they needed to sell anyway, so not many would be sold, so instead of making 0$ on the rest, it makes sense to sell those other cards. Because the fact is that Titan Vs aren't their main product, it's just a byproduct. Wait had to make sense is that they make a profit with the Tesla cards and that they make as much money as they can with the byproducts.

[Razor01]

2 minutes ago, laminutederire said:

If you want. You're not listening to people so I'm not going to reiterate after this time:

If you have two good gpus, with one you sell at 10 to 20k$, then the other at 3k$, you still make profit and most importantly: you sell something that you wouldn't sell otherwise because they have already sold most of the Tesla cards they needed to sell anyway, so not many would be sold, so instead of making 0$ on the rest, it makes sense to sell those other cards. Because the fact is that Titan Vs aren't their main product, it's just a byproduct. Wait had to make sense is that they make a profit with the Tesla cards and that they make as much money as they can with the byproducts.

 

 

No company will ever do that they will not take a loss a product that has no competition.  They will not bring that product to market it!

 

Getting 0 bucks is a loss because there are fixed costs associated with development and business that aren't being recouped, and those don't show up in margin figures.

[dalekphalm]

13 minutes ago, Razor01 said:

 

 

They will not sell a 3k card if they are making 0 bucks out of them, period.  They would not even release the card.  At 2 or 3 good GPU's per wafer, they will make no money.

 

The memory alone is 400 to 500 bucks man for them to buy.  Each stack of HBM 2, 4gb, is anywhere from 120 to 180 bucks.

Not necessarily. Remember that an already produced die, just sitting on a shelf, has already cost them money. If they have unused stock of V100 dies that they can't sell as Tesla Accelerators (let's say due to lack of demand for them), it would be in NVIDIA's best interest to sell them at loss, to recoup some of that initial investment.

 

Companies sell things at loss all the time - especially if just letting it sit there loses them more money than selling it at a loss does.

 

Especially if they have projections that yields will go up in the future, so the product stack will become profitable at a later date. Those initial dies already cost them, might as well sell them, especially if the next batch of stock might have yields enough to be profitable.

[Razor01]

8 minutes ago, dalekphalm said:

Not necessarily. Remember that an already produced die, just sitting on a shelf, has already cost them money. If they have unused stock of V100 dies that they can't sell as Tesla Accelerators (let's say due to lack of demand for them), it would be in NVIDIA's best interest to sell them at loss, to recoup some of that initial investment.

 

Companies sell things at loss all the time - especially if just letting it sit there loses them more money than selling it at a loss does.

 

Especially if they have projections that yields will go up in the future, so the product stack will become profitable at a later date. Those initial dies already cost them, might as well sell them, especially if the next batch of stock might have yields enough to be profitable.

 

 

Then why not hold on to them for quadros then?  They will cost 6k or more for top end quadros.  Even a cut down quadro 2nd one down will cost 4 to 5k.

[dalekphalm]

10 minutes ago, Razor01 said:

 

 

Then why not hold on to them for quadros then?  They will cost 6k or more for top end quadros.

There could be numerous reasons for that, including a lack of demand in the particular product segment you're talking about.

 

My point is that you're saying it's "impossible!" that NVIDIA might sell a die at a loss. We're showing you that this is quite simply, just not true.

 

Look, you might be right about the yields - that's not even the point. The point is that you're saying things as 100% definitive facts, then there's a lot of grey area and unclear information. Especially when the NVIDIA CEO says something that implies a different truth vs what you say.

 

It's entirely possible that NVIDIA, under the right circumstances, would sell V100 dies at a loss, to recoup some of the manufacturing cost. Is this what happened? Probably not. I don't know for certain either way. And neither do you.

[hobobobo]

20 minutes ago, Razor01 said:

 

 

Then why not hold on to them for quadros then?  They will cost 6k or more for top end quadros.  Even a cut down quadro 2nd one down will cost 4 to 5k.

The simplest reason would be to grab even more mindshare . Considering you can only get a Titan directly from nV, i doubt they sell that many of them and the loss on a single card bothers them at all. But the appearence of having a new beastly card buys them so much, speculation on the new 2080 (since the tradition of xx80ti being a titan but cheaper and later) alone is heaps of cash you dont have to spend in marketing.

[Razor01]

15 minutes ago, dalekphalm said:

There could be numerous reasons for that, including a lack of demand in the particular product segment you're talking about.

 

My point is that you're saying it's "impossible!" that NVIDIA might sell a die at a loss. We're showing you that this is quite simply, just not true.

 

Look, you might be right about the yields - that's not even the point. The point is that you're saying things as 100% definitive facts, then there's a lot of grey area and unclear information. Especially when the NVIDIA CEO says something that implies a different truth vs what you say.

 

It's entirely possible that NVIDIA, under the right circumstances, would sell V100 dies at a loss, to recoup some of the manufacturing cost. Is this what happened? Probably not. I don't know for certain either way. And neither do you.

 

1 minute ago, hobobobo said:

The simplest reason would be to grab even more mindshare . Considering you can only get a Titan directly from nV, i doubt they sell that many of them and the loss on a single card bothers them at all. But the appearence of having a new beastly card buys them so much, speculation on the new 2080 (since the tradition of xx80ti being a titan but cheaper and later) alone is heaps of cash you dont have to spend in marketing.

 

Its stupid to sell at a loss if they don't need to.

 

At this point there was no reason for a Volta card at 3k, the AI market is all nV's. 

 

They sold P100 as quadro's in last gen at 6k a pop, they can do the same thing now and still have those being sold as "low cost" AI development pieces.

 

We are talking about a company that always looks at the bottom line.  The last time they have ever sold at close to cost was with the tesla GT200 series.  That was because they were forced to cut down their profits due to competition.

[dalekphalm]

Just now, Razor01 said:

 

 

Its stupid to sell at a loss if they don't need to.

 

At this point there was no reason for a Volta card at 3k, the AI market is all nV's. 

 

They sold cut down P100 as quadro's in last gen at 6k a pop, they can do the same thing now and still have those being sold as "low cost" AI development pieces.

Again, you're either not seeing our arguments, or you're wilfully ignoring it.

 

All of what you said could be true. Just don't act like it's unchallenged fact. Unless you're executive level management in NVIDIA, you're just making assumptions. Those assumptions might be true - but assumptions they are.

 

One could also say "It's stupid to let this die we already paid for just sit on a shelf when we don't need to".

[Razor01]

4 minutes ago, dalekphalm said:

Again, you're either not seeing our arguments, or you're wilfully ignoring it.

 

All of what you said could be true. Just don't act like it's unchallenged fact. Unless you're executive level management in NVIDIA, you're just making assumptions. Those assumptions might be true - but assumptions they are.

 

One could also say "It's stupid to let this die we already paid for just sit on a shelf when we don't need to".

When was the last time nV sold at a loss or no profit for their products then?

 

I just edited the previous post to reflect that.  They were forced to do it last time.  They didn't even do it with the FX series ya know that?  You know why, that is business, they knew they were going to loose out with the FX series so they said screw it, its not worth going into a price war because they were going to get even less they can get anyways.  In this situation there is no advantage to marketing, or anything for that matter, to release a product that gives them no profits.  Because its already their market.

 

nV will never give us something for nothing.  Even Intel wouldn't do this, they aren't going to go down that road even when they get hammered tech wise.  Of course Intel did it behind closed doors, that is different though and illegal lol.

[hobobobo]

9 minutes ago, Razor01 said:

 

 

Its stupid to sell at a loss if they don't need to.

 

At this point there was no reason for a Volta card at 3k, the AI market is all nV's. 

 

They sold P100 as quadro's in last gen at 6k a pop, they can do the same thing now and still have those being sold as "low cost" AI development pieces.

Its really not though, especially if the production is not mature enough to pull the trigger on the new geforce lineup. Despite pascal still being the top dog, the fact that its 2y/o is hurting them. This way they have a new card to parade around as "the new king" and with the outlandish cost they insure that the loss they will take on it is well worth what they will gain in return. I mean, if they sold those dies to AIBs, it would be a completly different question, this way the are completly in control of stock and thus in control of their "losses". I mean, you look at a product as if the margin from the product is the only thing that matters, when its far from it. Investors need to be appeased (and again, pascal being 2yo begs the question if nV is getting complacent), gamers need to be whipped into a new frenzy with "omg look at those framerates" and AMD needs to be further shat on since they have nothing that holds a candle to Titan V. You can NOT achive that with gv100, or quadros for that matter, sales, its an invisible product to the broader market

[Razor01]

3 minutes ago, hobobobo said:

Its really not though, especially if the production is not mature enough to pull the trigger on the new geforce lineup. Despite pascal still being the top dog, the fact that its 2y/o is hurting them. This way they have a new card to parade around as "the new king" and with the outlandish cost they insure that the loss they will take on it is well worth what they will gain in return. I mean, if they sold those dies to AIBs, it would be a completly different question, this way the are completly in control of stock and thus in control of their "losses". I mean, you look at a product as if the margin from the product is the only thing that matters, when its far from it. Investors need to be appeased (and again, pascal being 2yo begs the question if nV is getting complacent), gamers need to be whipped into a new frenzy with "omg look at those framerates" and AMD needs to be further shat on since they have nothing that holds a candle to Titan V. You can NOT achive that with gv100 sales, its an invisible product to the broader market

 

 

Is a 2 year old product hurting them when they are making so much money on them,  Still breaking record quarters.  This product is not for mass gamers either, so its not going to be replacing anything.  And Q1 is a weaker than normal quarter but still broke records.

 

Again they could have released a Quadro p100 replacement and gotten more money and have the same affect you have stated.  That would appease the shareholder even more right?

 

Oh right now AMD doesn't hold a candle to GP102, so that affect is diminished anyways.