Nvidia's Pascal is Insane

[Skrillex24]

So I'm sitting here at work bored just looking around the Internet, and I came across Nvidia announcing some stuff about pascal at GTC.
16GB of HBM2 memory and 1TB of BANDWIDTH?
So happy I didn't upgrade my 770 to a 970 and gained 10 fps in games.
2016 is looking to be a great year

 

[Tom Waller]

Wonder what the price tag will be

 

[petek480]

It will probably be expensive. Considering nvidia cards that go for $700 are sold out for the first month at least there's really no reason for them to try to keep the price that low or offer a cheap version right away.

 

[wildfire707]

Have you seen the sample cards? Those PCBs are very complex to allow access to all of the memory simultaneously.

Anything using HBM style memory will be expensive due to the new design with added complexity of both the chips and the boards that run them.

 

[Mousemonkey]

@OP, watch your language.

 

[Skrillex24]

Lol don't tell me what I can and can't do

 

[Mousemonkey]

Oh but I can and will tell you what you can and cannot do on this forum.

 

[tracker45]

They are 10 per cent faster than current generation.

Not worth it in my opinion.

 

[Taylis_4]

They are 10 per cent faster than current generation.

Not worth it in my opinion.

I thought it was 20%? Whatever it is, it still hopefully should be the difference from around 50 fps to around 60 fps at 4K.

 

[Mousemonkey]

They are 10 per cent faster than current generation.

Not worth it in my opinion.

Where did you get that from? I don't care whether you're right or wrong but as the cards have yet to be released I'm curious as to your source.

 

[tracker45]

Their shareholders would never allow them to release products twice as fast as the previous generation.

They would lose way too much money from the older generations, and some people still buy the 600 series, 700 series and obviously the 900 series.

Their shareholders demand the new generation has an equivalent card of 10-20% increase.

Do you honestly think releasing something twice as fast as a gtx 960 (a 1060), at the same price is good for business ?

 

[looncraz]

Their shareholders would never allow them to release products twice as fast as the previous generation.

They would lose way too much money from the older generations, and some people still buy the 600 series, 700 series and obviously the 900 series.

Their shareholders demand the new generation has an equivalent card of 10-20% increase.

Do you honestly think releasing something twice as fast as a gtx 960 (a 1060), at the same price is good for business ?

That's why both nVidia and AMD will be releasing their middle-of-the-road solutions first.

Any super-fast variants will be limited to server/compute platforms for quite some time.

Once inventories and demand drops for the current generation, they will release the big dogs.

HBM2 isn't quite ready enough for mass debut anyway.

 

[bradum]

That's why both nVidia and AMD will be releasing their middle-of-the-road solutions first.

Any super-fast variants will be limited to server/compute platforms for quite some time.

Once inventories and demand drops for the current generation, they will release the big dogs.

HBM2 isn't quite ready enough for mass debut anyway.

Where are you getting this intel from?

Samsung and Hynix are only producing 4GB stacks of HMB2 right now, which maxes out at 16GB on one card. Since the Tesla workstation cards are going to feature 32GB HMB2, that means the Titan variant of the GP100 chip will debut first.

If it is not the Titan that is the first card to debut, it will be the GTX 1080 (or whatever it may be called) that is the first card to come out... Something I don't consider a "middle of the road" card. The x80 card has always come before the x60 and x50 cards.

 

[Zerk2012]

I love people assuming what will be released first, and how much of a increase it will be.

 

[bradum]

Their shareholders would never allow them to release products twice as fast as the previous generation.

They would lose way too much money from the older generations, and some people still buy the 600 series, 700 series and obviously the 900 series.

Their shareholders demand the new generation has an equivalent card of 10-20% increase.

I'd like to start off by saying: No, I don't think Pascal cards are going to be twice as fast as Maxwell cards.

With that being said... They will not be limited from 10 to 20% improvement by the shareholders.

Firstly, you are confusing the stock of certain retailers with the production of Nvidia cards. No card manufacturers are making 600 series cards any more, and haven't been for quite some time. The only 700 series card that may or many not still be in production is the 750ti, and I get the feeling the release of a high end card like the GTX 1080 or Titan would not affect the sales of people in the market for a 750ti. Nvidia is smart enough to stop production of a previous generation chip in anticipation of a new one coming out. This isn't to say that retailers will run out of stock immediately... The old cards will probably still be sold for another few months, but at a reduced price. Since all the chips have already been bought from Nvidia, they really don't care what price they sell at, and neither will their shareholders.

Secondly, If I'm an Nvidia shareholder, and Nvidia has the ability to produce a chip that will blow the AMD chips out of the water for a similar price, RELEASE THAT CHIP! They will sell incredibly well and Nvidia share prices will skyrocket.

 

[Zerk2012]

Why sell a card one time when you can get a decent amount of performance increase and the sell the same people a new card with a refresh model.

 

[looncraz]

Where are you getting this intel from?

Various sources, but especially Zauba manifests. And some logical deduction (HBM2 has only been in production for about a month), and large dies on 16FF+ or 14nm LPP are new territory (I don't think either process has ever mass-produced even a 200mm2 die). With ~0.4defects/mm2, yields aren't going to be that great for a 300mm2+ GPU. In addition, 16FF+ is only about twice as dense as 28nm, therefore a 300mm2 16FF+ die, which is the rumor for the Pascal parts seen in testing, is about the same as a 600mm2 die, so transistor counts can't be as high as nVidia or AMD have both claimed their flagship GPUs will possess.

Still, with both companies introducing either revised or nearly new architectures, we should see a healthy per-transistor performance increase. But beyond nVidia reintroducing dual-precision compute capabilities it stripped from Maxwell, I don't expect to see the full capabilities of this generation revealed until fall, maybe even winter.

I could always be wrong

If it is not the Titan that is the first card to debut, it will be the GTX 1080 (or whatever it may be called) that is the first card to come out... Something I don't consider a "middle of the road" card. The x80 card has always come before the x60 and x50 cards.

The first cards out may well beat the current top cards, but I doubt very seriously that we will see the top-performing cards from the start. With a mature process, that might be more likely. As this is such a giant leap in process technology and performance, it really isn't necessary for either company to take unnecessary risks. It could be that both companies have made such huge architectural improvements that they don't need the extra die space, but their claims of 18B+ transistors won't hold up in the rumored die space.

AMD's side is apparently further along towards release than nVidia, as we've seen many specific products being tested (via their code-names). Yet, we haven't seen a single shipping manifest regarding an HBM-compatible die. Everything has had the GDDR5 controller specifically stated in the manifest. Though, there's also an interesting fact that the manifests report the country of origin as Canada, which means these cards came out of AMD's graphics GQ (formerly ATi's HQ). That could mean that AMD is intentionally keeping their top-end SKUs under wraps by testing them in-house, or allowing it to be done in Taiwan or wherever... or that HBM2 is holding up the show (which would also impact nVidia).

 

[bradum]

Where are you getting this intel from?

Various sources, but especially Zauba manifests. And some logical deduction (HBM2 has only been in production for about a month), and large dies on 16FF+ or 14nm LPP are new territory (I don't think either process has ever mass-produced even a 200mm2 die). With ~0.4defects/mm2, yields aren't going to be that great for a 300mm2+ GPU. In addition, 16FF+ is only about twice as dense as 28nm, therefore a 300mm2 16FF+ die, which is the rumor for the Pascal parts seen in testing, is about the same as a 600mm2 die, so transistor counts can't be as high as nVidia or AMD have both claimed their flagship GPUs will possess.

Still, with both companies introducing either revised or nearly new architectures, we should see a healthy per-transistor performance increase. But beyond nVidia reintroducing dual-precision compute capabilities it stripped from Maxwell, I don't expect to see the full capabilities of this generation revealed until fall, maybe even winter.

I could always be wrong

If it is not the Titan that is the first card to debut, it will be the GTX 1080 (or whatever it may be called) that is the first card to come out... Something I don't consider a "middle of the road" card. The x80 card has always come before the x60 and x50 cards.

The first cards out may well beat the current top cards, but I doubt very seriously that we will see the top-performing cards from the start. With a mature process, that might be more likely. As this is such a giant leap in process technology and performance, it really isn't necessary for either company to take unnecessary risks. It could be that both companies have made such huge architectural improvements that they don't need the extra die space, but their claims of 18B+ transistors won't hold up in the rumored die space.

AMD's side is apparently further along towards release than nVidia, as we've seen many specific products being tested (via their code-names). Yet, we haven't seen a single shipping manifest regarding an HBM-compatible die. Everything has had the GDDR5 controller specifically stated in the manifest. Though, there's also an interesting fact that the manifests report the country of origin as Canada, which means these cards came out of AMD's graphics GQ (formerly ATi's HQ). That could mean that AMD is intentionally keeping their top-end SKUs under wraps by testing them in-house, or allowing it to be done in Taiwan or wherever... or that HBM2 is holding up the show (which would also impact nVidia).

Man, look at the last manufacturing process change, and the one before that... Never has Nvidia come out with a mainstream consumer card before their top offerings.
And it's not just Nvidia... Never, when a tech company is coming out with a new line of products, do they start with the budget consumer version first.

All that we know for sure is that pascal cards are in testing, and HBM2 MASS production began a month ago... The rest is simply speculation by you.

 

[looncraz]

Man, look at the last manufacturing process change, and the one before that... Never has Nvidia come out with a mainstream consumer card before their top offerings.
And it's not just Nvidia... Never, when a tech company is coming out with a new line of products, do they start with the budget consumer version first.

All that we know for sure is that pascal cards are in testing, and HBM2 MASS production began a month ago... The rest is simply speculation by you.

Actually, nVidia released the mainstream chips for GTX 900 series first GTX 970 & 980 on 9/18/14. They then released the mobile lineup in stages. They only released the Titan X six months after the mainstream parts - on 3/17/15. Another six weeks passed before we saw the 980Ti (6/2/15).

They didn't do the same thing with the GTX700 series, with the Titan being released 2/19/13 and the more mainstream parts coming a month later. The GTX680 came out before the 670 (two months), the GTX580 a month before the 570...

But it all comes down to if the high end is using a larger die than the mainstream. The GTX900 series had that quality (as will Polaris, most likely) and Pascal undoubtedly will. Rumors have the big chips with 18B transistors - more than double needed for mainstream performance cards.

It's important to remember that this is the first foray for both companies using what are, effectively, mobile-tuned processes that have ever only been used for 100mm2 or so dies and have only 70% yields with those. Large dies will have much worse yields, so going to the larger dies will be slower. In addition, both companies' high end cards will be using HBM2 - a technology which only entered into mass production a month ago - the first batch is likely not even completed, but should be weeks away, which adds another level of delay that only applies to the high end cards... including the interposer imposed delay.

Now, it's possible we may see Tesla compute GPUs first, no doubt. But indications are favoring a mainstream, GDDR5-based, launch occurring first - just like nVidia did with the GTX900 series.

 

[renz496]

Where are you getting this intel from?

Various sources, but especially Zauba manifests. And some logical deduction (HBM2 has only been in production for about a month), and large dies on 16FF+ or 14nm LPP are new territory (I don't think either process has ever mass-produced even a 200mm2 die). With ~0.4defects/mm2, yields aren't going to be that great for a 300mm2+ GPU. In addition, 16FF+ is only about twice as dense as 28nm, therefore a 300mm2 16FF+ die, which is the rumor for the Pascal parts seen in testing, is about the same as a 600mm2 die, so transistor counts can't be as high as nVidia or AMD have both claimed their flagship GPUs will possess.

Still, with both companies introducing either revised or nearly new architectures, we should see a healthy per-transistor performance increase. But beyond nVidia reintroducing dual-precision compute capabilities it stripped from Maxwell, I don't expect to see the full capabilities of this generation revealed until fall, maybe even winter.

I could always be wrong

If it is not the Titan that is the first card to debut, it will be the GTX 1080 (or whatever it may be called) that is the first card to come out... Something I don't consider a "middle of the road" card. The x80 card has always come before the x60 and x50 cards.

The first cards out may well beat the current top cards, but I doubt very seriously that we will see the top-performing cards from the start. With a mature process, that might be more likely. As this is such a giant leap in process technology and performance, it really isn't necessary for either company to take unnecessary risks. It could be that both companies have made such huge architectural improvements that they don't need the extra die space, but their claims of 18B+ transistors won't hold up in the rumored die space.

AMD's side is apparently further along towards release than nVidia, as we've seen many specific products being tested (via their code-names). Yet, we haven't seen a single shipping manifest regarding an HBM-compatible die. Everything has had the GDDR5 controller specifically stated in the manifest. Though, there's also an interesting fact that the manifests report the country of origin as Canada, which means these cards came out of AMD's graphics GQ (formerly ATi's HQ). That could mean that AMD is intentionally keeping their top-end SKUs under wraps by testing them in-house, or allowing it to be done in Taiwan or wherever... or that HBM2 is holding up the show (which would also impact nVidia).

Man, look at the last manufacturing process change, and the one before that... Never has Nvidia come out with a mainstream consumer card before their top offerings.
And it's not just Nvidia... Never, when a tech company is coming out with a new line of products, do they start with the budget consumer version first.

All that we know for sure is that pascal cards are in testing, and HBM2 MASS production began a month ago... The rest is simply speculation by you.

that is if you only look at the graphic card names. but for some people they take those GK104 (GTX680) and GM204 (GTX980) as a 'mainstream' chip.

 

[bradum]

Where are you getting this intel from?

Various sources, but especially Zauba manifests. And some logical deduction (HBM2 has only been in production for about a month), and large dies on 16FF+ or 14nm LPP are new territory (I don't think either process has ever mass-produced even a 200mm2 die). With ~0.4defects/mm2, yields aren't going to be that great for a 300mm2+ GPU. In addition, 16FF+ is only about twice as dense as 28nm, therefore a 300mm2 16FF+ die, which is the rumor for the Pascal parts seen in testing, is about the same as a 600mm2 die, so transistor counts can't be as high as nVidia or AMD have both claimed their flagship GPUs will possess.

Still, with both companies introducing either revised or nearly new architectures, we should see a healthy per-transistor performance increase. But beyond nVidia reintroducing dual-precision compute capabilities it stripped from Maxwell, I don't expect to see the full capabilities of this generation revealed until fall, maybe even winter.

I could always be wrong

If it is not the Titan that is the first card to debut, it will be the GTX 1080 (or whatever it may be called) that is the first card to come out... Something I don't consider a "middle of the road" card. The x80 card has always come before the x60 and x50 cards.

The first cards out may well beat the current top cards, but I doubt very seriously that we will see the top-performing cards from the start. With a mature process, that might be more likely. As this is such a giant leap in process technology and performance, it really isn't necessary for either company to take unnecessary risks. It could be that both companies have made such huge architectural improvements that they don't need the extra die space, but their claims of 18B+ transistors won't hold up in the rumored die space.

AMD's side is apparently further along towards release than nVidia, as we've seen many specific products being tested (via their code-names). Yet, we haven't seen a single shipping manifest regarding an HBM-compatible die. Everything has had the GDDR5 controller specifically stated in the manifest. Though, there's also an interesting fact that the manifests report the country of origin as Canada, which means these cards came out of AMD's graphics GQ (formerly ATi's HQ). That could mean that AMD is intentionally keeping their top-end SKUs under wraps by testing them in-house, or allowing it to be done in Taiwan or wherever... or that HBM2 is holding up the show (which would also impact nVidia).

Man, look at the last manufacturing process change, and the one before that... Never has Nvidia come out with a mainstream consumer card before their top offerings.
And it's not just Nvidia... Never, when a tech company is coming out with a new line of products, do they start with the budget consumer version first.

All that we know for sure is that pascal cards are in testing, and HBM2 MASS production began a month ago... The rest is simply speculation by you.

that is if you only look at the graphic card names. but for some people they take those GK104 (GTX680) and GM204 (GTX980) as a 'mainstream' chip.

He originally said mid-range. The 980 and 970 are not mid-range.

 

[bradum]

Actually, nVidia released the mainstream chips for GTX 900 series first GTX 970 & 980 on 9/18/14. They then released the mobile lineup in stages. They only released the Titan X six months after the mainstream parts - on 3/17/15. Another six weeks passed before we saw the 980Ti (6/2/15).

They didn't do the same thing with the GTX700 series, with the Titan being released 2/19/13 and the more mainstream parts coming a month later. The GTX680 came out before the 670 (two months), the GTX580 a month before the 570...

But it all comes down to if the high end is using a larger die than the mainstream. The GTX900 series had that quality (as will Polaris, most likely) and Pascal undoubtedly will. Rumors have the big chips with 18B transistors - more than double needed for mainstream performance cards.

It's important to remember that this is the first foray for both companies using what are, effectively, mobile-tuned processes that have ever only been used for 100mm2 or so dies and have only 70% yields with those. Large dies will have much worse yields, so going to the larger dies will be slower. In addition, both companies' high end cards will be using HBM2 - a technology which only entered into mass production a month ago - the first batch is likely not even completed, but should be weeks away, which adds another level of delay that only applies to the high end cards... including the interposer imposed delay.

Now, it's possible we may see Tesla compute GPUs first, no doubt. But indications are favoring a mainstream, GDDR5-based, launch occurring first - just like nVidia did with the GTX900 series.

Again, no one would refer to something like a 1080 as "Mid-range". And if you read my responses from before, you'd see I said it would be EITHER the Titan or the 1080.

I can promise you they're not coming out with cards like the 1050 and 1060 first.

 

[looncraz]

Again, no one would refer to something like a 1080 as "Mid-range". And if you read my responses from before, you'd see I said it would be EITHER the Titan or the 1080.

I can promise you they're not coming out with cards like the 1050 and 1060 first.

Just a conflict of terminology, then ;-)

To me, the GTX 970 is the epitome of the mainstream card - considering it is the best-selling card next to the 750Ti, I think that definitely counts.

http://www.amazon.com/Best-Sellers-Electronics-Computer-Graphics-Cards/zgbs/electronics/284822

The 980, though, isn't really mainstream, it's entry-level high-end. Either way, these are the types of cards that will be out first. They will likely match/slightly outperform the Titan X/980Ti in performance, but represent the new class of GPU. The same thing on AMD's side is also likely for reasons I've already mentioned.

 

[renz496]

It depends on how you want to interpret them. But for those that follow gpu tech for so long those chip inside 680 and 980 are considered as mid range chip. They were positioned as high end because of relative performance to competitor high end gpu performance at the time of release. Up to 500 series chip with code name ended with number 4 has always been nvidia mid range offering. If AMD really competitive with their performance you will see those GM204 debut as GTX960 series instead of GTX980/970.