NVIDIA Pascal Thread

[DooKey]

ehm no,

Mainstream used to be at $250, now its $330-500

Example, GTX560 vs GTX 680 and GTX980

High-End was used to be $500, now its $750-1000

Example, GTX 580 vs GTX 780Ti/Titan and GTX 980Ti/Titan

edit: And just to understand exactly what has happened.

GTX 560Ti = GF114 = 332mm2 = MSRP $250
GTX 680 = GK104 = 294mm2 = MSRP $500
GTX980 = GM204 = 398mm2 = MSRP $549

GTX 580 = GF110 = 520mm2 = MSRP $500
GTX 780Ti = GK110 = 560mm2 = MSRP $700
GTX 980Ti = GM200 = 600mm2 = MSRP $650
TITAN-X = GM200 = 600mm2 = MSRP $1000

You can't live in the past. The market has changed and will continue to do so. Nodes are increasingly expensive to develop and the market for dGPU is shrinking. Consumer prices are going to go up.

 

[coercitiv]

You can't live in the past. The market has changed and will continue to do so. Nodes are increasingly expensive to develop and the market for dGPU is shrinking. Consumer prices are going to go up.

Are GPU companies making less money from gaming products nowadays?

 

[tviceman]

Sweepr,

This is from Micron's most recent earnings call:



Micron's third fiscal quarter ends in early June. Could see 1070/1080 w/ GDDR5 in June, and then 1080 Ti with GDDR5X in August/September.

Close to my thoughts. I think the 1080 will be a GP104 with "slower" 10gb/s GDDR5X and the 1080 TI will still be a GP104, but with higher clocks and 12-14gb/s GDDR5X. GDDR5 for the 1070. $699-749, $549-599, and $399 all for GP104 based cards.

 

[Arachnotronic]

Look at the die sizes of the chips and you will understand what we are saying.

This is not a compelling argument.

Die sizes alone don't tell you much. The number of transistors that can be packed into a given area has gone up substantially, so has the performance of said transistors. The cost, though, has been higher cost per unit area and additional complexity in terms of design.

A 600mm^ part on 16FF+ will cost substantially more to build than a 600mm^2 part on 28nm to build.

 

[Kris194]

You can't live in the past. The market has changed and will continue to do so. Nodes are increasingly expensive to develop and the market for dGPU is shrinking. Consumer prices are going to go up.

Prices going up is one thing, lying consumers is the second.

This is not a compelling argument.

Die sizes alone don't tell you much. The number of transistors that can be packed into a given area has gone up substantially, so has the performance of said transistors. The cost, though, has been higher cost per unit area and additional complexity in terms of design.

A 600mm^ part on 16FF+ will cost substantially more to build than a 600mm^2 part on 28nm to build.

Ah so let's say that we shouldn't pay 100K for cars but 200K because they are faster? Oh wait, things getting better is something what we call progress.

Look at this, cards from the same range

2010

GF104 - GTX 460

2014

GM204 - GTX 980

 

[DooKey]

Are GPU companies making less money from gaming products nowadays?

Those that do will drop out of the gaming market eventually.

 

[Arachnotronic]

Kris194, still haven't made a good argument in defense of your position.

 

[coercitiv]

Those that do will drop out of the gaming market eventually.

Let me rephrase that: is Nvidia getting lower margins on gaming chips than it did years ago?

 

[DooKey]

Let me rephrase that: is Nvidia getting lower margins on gaming chips than it did years ago?

Not that I'm aware of. Are you saying they should be?

 

[KaRLiToS]

This is not a compelling argument.

Die sizes alone don't tell you much. The number of transistors that can be packed into a given area has gone up substantially, so has the performance of said transistors. The cost, though, has been higher cost per unit area and additional complexity in terms of design.

A 600mm^ part on 16FF+ will cost substantially more to build than a 600mm^2 part on 28nm to build.

Do you have any source for the cost difference? Show me numbers instead. You sound like you appreciate to pay more for less.

 

[coercitiv]

Not that I'm aware of. Are you saying they should be?

With a shrinking market and ever more expensive nodes? Of course they should.

 

[jpiniero]

Do you have any source for the cost difference? Show me numbers instead. You sound like you appreciate to pay more for less.

Well, there's this...

AMD badly needs money so don't expect them to engage in a price war.

 

[CakeMonster]

That being said, I don't think the GTX 1070 will be cheap. New node (FinFET is more expensive), strained world economy, currency wars between China and the US, weak consumer spending, inflation etc. all add up to the price.

Isn't that why they're pushing small, simple to produce, mid-range chips as high-end cards? They didn't do that before and I would assume that more than makes up for the margins lost to the world economy situation.

 

[RussianSensation]

Let me rephrase that: is Nvidia getting lower margins on gaming chips than it did years ago?

NV has set record gross margins and record GPU profits for every consecutive new GPU generation starting with Fermi. If you read their financials from 2009 to 2016:

- average selling prices per SKU are WAY up
- Gross profit margins are WAY WAY up
- Net income is WAY up

That means the idea that it costs NV exponentially more to manufacture chips on new nodes is not reality. Sure, it might cost more initially but NV keeps prices fixed for the duration of their cards useful life until they are replaced or discontinued.

I said in 2012 that once NV sets new price levels and consumers accept, that is it. Moving forward, it will be accepted as the status quo. They did it again with Maxwell and once again they bifurcated a generation.

Ok fine, the market has changed and I accept it since more consumers voted against those of us who opposed and skipped the purchase. But then let's call a spade a spade. If we accept the new market, let's accept the truth then -- we are being asked to pay $550-600 for mid-range products of a next gen product line. People saying die sizes don't matter? They do matter since they tell us where the product SKU sits in the entire stack of next generation series; just like it mattered before.

Like I said, if NV artificially creates a 20-25% gap between a 1070 and 1080, that 1070 card is only a marketing name. Why are people already defending this?

Like I said, you guys want to accept the truth and that the market has changed, then man up and accept it. If 1080 costs $600-650 and it beats 1070 by 20-25%, and NV prices 1070 at $349-399, it will mean exactly what I said -- NV changing the name of 1060Ti to 1070 while raising the price from $299 660Ti. For 1080, it would mean yet another price increase from $499 680 which was already 2X the price increase of a spiritual 560Ti predecessor.

I have no problem accepting that I won't pay $550-650 for upper mid-range products and I won't, but at least I don't defend this new direction in GPU pricing.

Here is a good example. Sennheiser $650 HD650 flagship was replaced by HD800 at $1500, upper mid-range HD700 now moved up to $1000! Now HD800 is replaced by HD800S for $1700. Let me tell you what happened in the Hi-Fi space -- not only did people defend these prices, but it became a sort of bragging rights to show off who has high-end gear. I guarantee it that if HD800S cost $999 (i.e., tech should get cheaper over time), it wouldn't be as desirable. Irony. It sounds awesome for the 3% of income earners until prices of the entire industry go up at every level.

Guess what's happening in GPUs now? The same thing. Defending, denials, and justification why these prices are justifiable.

Why don't we go back in time and rename GTX570 to GTX570Ti and jack it up to $649? Poor NV gave them away for $349.

Like I said, anyone is free to buy what they want but at least be a man and accept that we are now paying $500-550 for upper mid-range product line. I am not defending AMD either as I pointed out how they magically jacked up 290/290X's market prices by $100+ (per AnandTech's review).

Also, 680 was barely 9-10% faster than 670. If 1080 beats 1070 by 20-25%, yes, I will call out the marketing BS of 1070 name and how NV raised prices again should 1080 cost above $550. Just don't get offended about the truth.

Also, if NV gives 1070 GDDR5 and 1080 gets GDDR5X, that's BS if you ask me because most consumers won't read the specs and assume it's almost the same thing like they got screwed into buying a "4GB GDDR5 GTX970".

 

[CakeMonster]

Another possibility is lower power usage of 1070/1080 cards to leave headroom for their refreshes in 2017. However, I don't buy that since NV set TDP arbitrarily on 970/980 anyway and no one cared.

What do you mean by that? Arbitrarily like low? They never released a faster clocked version of either 970/980 AFAIK.

 

[Kris194]

Look what they did with Kepler.

GTX 670 - cut GK104
GTX 680 - full GK104

GTX 770 - full, slightly overclocked GK104
GTX 780 - cut GK110
GTX 780 Ti - full GK110

I believe that they will do the same thing with Pascal. If the chiphell is right then it will look like this

GTX 1070 - cut GP104, GDDR5
GTX 1080 - full GP104, GDDR5X

GTX 1170 - full GP104, GDDR5X, slightly overclocked
GTX 1180 - cut GP102

With a shrinking market and ever more expensive nodes? Of course they should.

Maybe nodes are more expensive but with new node you can make the same chip for much lower price, not to mention that you can make more chips of the same wafer. Like I said, I have nothing against prices going up but I do not agree for lying consumers. Look how GTX 460 1GB performed when it was released and now look how GTX 980 performed when it was released in most popular resolutions. Do you see any similarities? I do.

 

[RussianSensation]

Well, there's this...

AMD badly needs money so don't expect them to engage in a price war.

Polaris 10/11 are likely to target different markets than 1070/1080.

As far as the cost argument goes, NV marketing uses this as a crutch for VRAM gimping Fermi, Kepler and Maxwell stacks. Why is it NV has tried to sell bare minimum VRAM while charging absurd premiums on GTX570 2.5GB, GTX580 3GB, GTX680/770 4GB, GTX780 6GB.

AMD launched a 512-bit bus (complex PCB), 275W (complex VRM/power circuitry and expensive cooling) with 8GB for $329. That's about 1 year ago from the time 1070/1080 drop. I bet 256-bit 8GB GDDD5X isn't that expensive in comparison and GP104 will not have a 438mm2 die size. So take higher costs of 14nm/16nm I bet most of that is offset by these other aspects. The real reason NV won't go back to charging $299-329 for mid-range 1080 isn't cost, but because consumers keep paying. That's the truth. If you go back to 2009-2012 NV financials, their gross margins were never in the 53-56% range. Go check. This is key since its Revenue - COGS (manufacturing included). That means NV is more than compensating for whatever higher costs of newer nodes with higher ASP.

I guess you could make the argument that if NV went back to historical prices and old gross margins, then AMD would go bankrupt. Fine, then at least don't do stupid marketing tricks like VRAM gimping and creating massive artificial performance gap between 1080 and 1070. At least if you are going to charge high prices, then deliver the appropriate performance.

If 1080 outperforms 1070 by 23-26%, I will call out the BS that 1070 is really a 1060Ti, renamed 1070.

BTW, rumor still has it that Pascal hasn't incorporated proper Async Compute support:
http://www.redgamingtech.com/asynchronous-compute-trouble-for-nvidias-pascal-architecture-rumor/

 

[jpiniero]

They have definitely raised prices at a given tier, no question. With costs now spiking with this node, don't expect things to change.

 

[Kris194]

Kris194, still haven't made a good argument in defense of your position.

Look how games performed on both cards in most popular resolution when they were released. If it's not enough for you then look, chips from the same range

July 2010 (40nm)
GTX 460 = GF114 = 332 mm² --> 229$

January 2011 (40nm)
GTX 560 Ti = GF114 = 332 mm² --> 249$

March 2012 (28nm)
GTX 680 = GK104 = 294 mm² --> 499$(sic!)

May 2013 (28nm)
GTX 770 = GK104 = 294 mm² --> 399$

September 2014 (28nm)
GTX 980 = GM204 = 398 mm² --> 549$

So to sum up, you're paying more for less.

 

[mohit9206]

Hmm not sure the hype is worth it. The way Pascal and Polaris have been hyped for so long expecting huge performance jump but i feel instead we will get incremental jump with the focus being not on performance but instead on performance per watt. Enthusiasts are gonna be disappointed. People are expecting 980ti performance at $400 from Pascal let's see if it happens or not.

 

[Qwertilot]

Well, the really huge jump vs the 980ti will have to wait for mainstream HBM2/huge chips. Due in a bit from both people, but not yet.

What NV will want - and P100 suggests they'll be able to do - is to get the 1080 enough above the 980ti to get the compulsive upgraders to bite. Won't need to be such a huge jump huge I guess, but likely to be present.

Price wise, well, if they do that they'll have a monopoly position on this market sector until Vega so.....

 

[CakeMonster]

What NV will want - and P100 suggests they'll be able to do - is to get the 1080 enough above the 980ti to get the compulsive upgraders to bite. Won't need to be such a huge jump huge I guess, but likely to be present.

Heh, that's partially my fear. I have a 980 which I'm very happy with and 980Ti was never an alternative. But I will definitely be tempted now. It could be a good buy, or it could be a bad one.. depending on when the big chip hits and how its performance compares.

 

[96Firebird]

Guess how prices go down?

Competition.

 

[flopper]

Guess how prices go down?

Competition.

Unless they talk about keeping them up behind closed doors.
wont expect anything entusiast end to be priced sensibly.

 

[JDG1980]

Well, there's this...

We shouldn't take this slide at face value. First of all, it's years old - it was first published on a website way back in March 2012. We aren't talking about current data, we're talking about 4-year-old rough estimates. Secondly, this slide was created by Nvidia as part of a presentation to put pressure on TSMC to lower prices. In other words, it's propaganda, not an objective and dispassionate analysis.

One of the major issues raised by Nvidia - having to pay premiums for early risk production - is now a complete nonissue because Apple now fills that role. That slide was written in a different era, one in which mobile devices were far less important and played a far smaller role in the development of new nodes.