nVidia Q3: Revenue up 53%

[Headfoot]

Next time when I park between 2 cars I'll tell people I parked "at the enthusiast level" between the cars. Since apparently the word "Middle" is verbotten.

Why is this so hard to accept? It's the middle card. That's reality. I don't understand why people are fighting this so hard. Its an accurate characterization by every metric.

 

[happy medium]

Your argument is "Here's some handwaving and my outrage"

My argument is look at these firm numbers and determine which one is in the middle. It's GP104, every time.

show me firm links, like the rules of Anandtech forum ask for, and I'll be happy.
Firm opinions don't count.

 

[b-mac]

How is pin connector in anyway equivalent to die size. You can keep ignoring all the info you want and keep parroting your GPU name argument.

 

[Headfoot]

show me firm links, like the rules of Anandtech forum ask for, and I'll be happy.
Firm opinions don't count.

What are you talking about? Did you not see the numbers in that list? Those are all accurate per nVidia's published specs.

Are you trying to tell me you think the 1080 doesn't have 8GB of GDDR5X and 320GB/s bandwidth and 2560 Cuda Cores?

 

[happy medium]

woops

 

[Headfoot]

show me firm links, like the rules of Anandtech forum ask for, and I'll be happy.
Firm opinions don't count.

mid·dle
ˈmidl/
adjective
adjective: middle

1. at an equal distance from the extremities of something; central.
   "the early and middle part of life"
   synonyms: central, mid, mean, medium, medial, median, midway, halfway
   "the middle point"
2. (of a member of a group, series, or sequence) so placed as to have the same number of members on each side."the woman was in her middle forties"
3. intermediate in rank, quality, or ability.
   • "there is a dearth of talent at the middle level"
     synonyms: intermediate, intermediary
     "the middle level"

source: https://www.google.com/webhp?source...=2&ie=UTF-8#q=dictionary+definition+of+middle

 

[happy medium]

How is pin connector in anyway equivalent to die size. You can keep ignoring all the info you want and keep parroting your GPU name argument.

the same way die size is the equivalent to tier placement.
I like information but I like links and fact to back them first.

like this..
I say these are the proper tiers of price performance.
I looked up the prices and performance for all cards in the past 7 years.

Super high end.
gtx 590 700$ March 2011
gtx 690 1000$ May 2012
Titan 1000$ March 2013,/ gtx780ti November 2013 ,700$
Titan X 1000$ March 2015, /gtx980ti June 2015 ,700$
Titan XP 1200$ August 2016

High end
gtx580 500$, November 2010
gtx680 500$ March 2012
gtx780 650$ June 2013
gtx980 550$ September 2014
gtx1080 650$ May 2016

lower high end
gtx570 350$ December 2010
gtx670 400$ May 2012
gtx770 400$ May 2013
gtx970 330$ September 2014 ,WOW! cheap
gtx1070 400$ June 2016

Mid range
gtx560ti 250$ January 2011
gtx660ti 300$ August 2012
gtx760 250$ June 2013
gtx960 220$ Jan 2015
gtx1060 250$ Aug 2016

lower end
gtx560
gtx660
gtx750
gtx950
gtx1050

then I looked for links that back upmy claims or my tiers of price/ performance,like this.

quote from gtx680 review "retaking the performance crown"

"Last but not least, let’s talk about pricing and availability. GTX 680 is the successor to GTX 580 and NVIDIA will be pricing it accordingly, with an MSRP of $500. This is the same price that the GTX 580 and GTX 480 launched at back in 2010, and while it’s consistent for an x80 video card it’s effectively a conservative price given GK104’s die size."
http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review

quote from the gtx780 review "The new high end"

"Meanwhile, compared to the GTX 680 which it will be supplanting, the GTX 780 should be a big step up in virtually every way. As NVIDIA likes to put it, GTX 780 is 50% more of everything than GTX 680. 50% more SMXes, 50% more ROPs, 50% more RAM, and 50% more memory bandwidth. In reality due to the clockspeed differences the theoretical performance difference isn’t nearly as large – we’re looking at just a 29% increase in shading/texturing/ROP performance – but this still leaves GTX 780 as being much more powerful than its predecessor."
http://www.anandtech.com/show/6973/nvidia-geforce-gtx-780-review

quote from the gtx980 review.

"Today’s launch will see GM204 placed into two video cards, the GeForce GTX 980 and GeForce GTX 970. We’ll dive into the specs of each in a bit, but from an NVIDIA product standpoint these two parts are the immediate successors to the GTX 780/780Ti and GTX 770 respectively.
http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review


quote from a gtx1080 preview.

"The base clockspeed of 1607MHz is some 42% higher than GTX 980 (and 60% higher than GTX 980 Ti), and the 1733MHz boost clockspeed is a similar gain. On paper, GTX 1080 is set to offer 78% better performance than GTX 980,"
http://www.anandtech.com/show/10326/the-nvidia-geforce-gtx-1080-preview/


I'm having trouble finding any information about a 1080 being the successor to the gtx980 because of its die size, power consumption, ram amounts.

This seems to be a myth brought to you by uniformed posters.

 

[Headfoot]

I don't understand why mid range is such a scary word... I pretty much buy almost everything in the mid range. It's almost always the best combination of performance for the price, in nearly every category of consumer goods. I bought a mid range car, a mid range house, a mid range snoblower....

 

[witeken]

Die size alone should tell you that the GTX 1080 is only a "mid end" card.
The full GP104, aka GTX 1080 is only 314mm² for godsake..

100mm² -> 175mm² = low end
176mm² -> 350mm² = mid end (GTX 1080)
351mm² -> 500mm² = high end
501mm² -> 600+mm² = ultra high end

~ numbers above

You realize that the cost per mm² is not constant. Right?

16nm is easily 30% more expensive per mm² to manufacture, so that makes your comparisons (28-16nm) not apples to apples.

 

[happy medium]

You realize that the cost per mm² is not constant. Right?

16nm is easily 30% more expensive per mm², so that makes your comparisons not apples to apples.

I was gonna use this argument and although correct might confuse people.

 

[Det0x]

You realize that the cost per mm² is not constant. Right?

16nm is easily 30% more expensive per mm², so that makes your comparisons not apples to apples.

Instead of partially quote my post, you should post the whole thing to put it into context.

Those numbers at the end of my post was only +- ballpark numbers to put the rest of my post into perspective, hence the "~".

 

[witeken]

I was gonna use this argument and although correct might confuse people.

Yeah, scrolling through the comments people seem to prefer simple arguments like "that GPU is midrange" and "no it is high-end". That, people, is semantics. Welcome in the happy world of linguistics.

 

[witeken]

Instead of partially quote my post, you should post the whole thing to put it into context.

Those numbers at the end of my post was only +- ballpark numbers to put the rest of my post into perspective, hence the "~".

My point is just that how many mm² is midrange or high-end is not a fixed number since the cost per mm² is not fixed in the first place.

 

[USER8000]

You realize that the cost per mm² is not constant. Right?

16nm is easily 30% more expensive per mm², so that makes your comparisons not apples to apples.

It affects everybody who makes chips for any product. ASP has gone up significantly more than any production cost increases for Nvidia graphics carrds.

If you even look at the top end 2010 iPad SKUs - it was $829 for a 64GB 10" model. Fast forward to 2016,the price has increased for to $1029 for the 10" iPad Pro and $1229 for the 13" model. There is a price increase,but it pales in comparison to graphics cards.

Using your excuses we should be seeing laptops costing 2 to 4 times more than 6 years ago and consoles too.

The GTX580 was $500 at launch and used a full 529MM2 GF114 chip and the Titan X Pascal is $1200 and uses a partially disabled 471MM2 GP102 chip . Since the Titan X Pascal does not use a fully enabled chip,it is closer to a GTX570 in some ways,which was $330 in 2010.

That means in reality,we have gone from $330 to $1200.

Yet,we have not seen these magical price 2 to 4 times price increases for all the other consumer products using the same TSMC/Samsung and Global Foundries 16NM/14NM process nodes.

If that was the case every company using them would be loosing big money on their next generation chips in phones and so on.

The XBox One S uses the same TSMC 16NM process as most of the Pascal cards.

I find it hard Microsoft would do a shrink which cost them more money.

 

[Pariah]

Market conditions and Nvidia's business strategy in Q3 were setup about as perfectly as they could be for Nvidia and the financial results speak for themselves.

For whatever reason, people here hold Nvidia to a different standard than other companies. Almost everything costs more than just a few years ago. Semiconductor compaies used to be able to offset inflation with die shrinks. As I posted earlier and all the AMD people ignored, that didn't happen with this node transition. Every other industry raises prices to compensate for increased costs, but Nvidia is supposed to be a charity for you people and eat the increased costs.

Funny how when Nvidia initially released their Pascal gpus and supply was scarce for weeks the mob here ripped Nvidia for rushing the launch to beat AMD and that it was a paper launch. Yea... so much for that theory. You don't produce the balance sheet Nvidia did in q3 by having only superficial stock available for what amounts to a paper launch.

Demand was obviously extreme when taking into account the early supply shortages despite the significant price gouging early on by retailers which Nvidia benefitted nothing from. Based on average selling prices Nvidia could have set the MSRP much higher than they did. You think if Nvidia set the msrp for the 1080 at $350 that any retailer would have sold it for that price? They still would have sold for $750+ because that's what the market determined they were worth. Yet on these boards Nvidia is ripping us off and price gouging us because they would rather increase their revenue instead of letting retailers collect all the profits.

 

[linkgoron]

I was gonna use this argument and although correct might confuse people.

The x80 card used to be the top-of-the-stack flagship single card, with a price of 500-600. Since Kepler, the 80 card has not been the top card, but its price has only increased. That's why people were saying that it's a fake flagship. The code-name changing is also revealing, as are the die sizes. Those are all characteristics of what used to be the mid-range of the stack. Smaller dies, and a X4 code name.

Unsurprisingly, given those characteristics, the 980 and 1080 really are still in the mid-range of the stack. The Titan X/980ti were the top vs the 980 and the Titan XP is the top of the stack vs the 1080. People guessed that the cards are too small to be the flagships, (see AT review quote) and those guesses were true, as die sizes actually have some technical meaning. The only thing that has changed is their name. Now, the mid range of the stack has a price of 500$-650$ an is called 1080, and the high-end of the stack has a price of 1200$ and is called Titan/x80Ti, while costing twice as much as it used to. What are you arguing about?

 

[kraatus77]

GX104 is always midrange because there's 1 more gpu class sits above and that is GX100 (and GX102

it's a fact, according to nvidia itself. no matter how much one tries and deny it. spamming same thing doesn't change a fact just because it hurts for some weird reason.

never seen a consumer fighting for paying more, it's absurd.

 

[Det0x]

My point is just that how many mm² is midrange or high-end is not a fixed number since the cost per mm² is not fixed in the first place.

Die size alone should tell you that the GTX 1080 is only a "mid end" card.
The full GP104, aka GTX 1080 is only 314mm² for godsake..

1000 series
The Pascal GTX 1050 TI is 135mm² (low end)
The Pascal GTX 1060 is 200mm² (mid end)
The Pascal GTX 1080 is 314mm² (mid end)
The Pascal "GTX 1080TI" / Titan X Pascal is 471mm² (high end)
The Pascal Tesla P100 is 610mm² (ultra high end)

Die size means everything when it comes to how much is BOM cost and how much it sells for. because that's where your huge pile of money /revenue/margins came from.

"if Nvidia made a gtx1080 with a 200mm dies size, with performance 2x faster than a gtx980ti, and sold it for 1000$. It would be a mid range card?

yes it will, it will be a mid range infront of 450-500mm die size gpu of same architecture. because it will be 2x faster than midrange gtx1080@200mm.

your argument is baseless and anti consumer at best.

Card1 @ 150mm2 is 1.5x faster then X-card (low end)
Card2 @ 200mm2 is 2x faster then X-card (mid end)
Card3 @ 450mm2 is 4.5x faster then X-card (high end)
Card4 @ 600mm2 is 6x faster then X-card (ultra high end)

Even if card2 is 2x faster!!111 then X-card, its still only mid end in the product stack

No, it's mid range. Take literally any measurable metric, let's take CUDA cores for example.

GP100 (DGX-1) -> 3840 Cores
GP102 (Titan XP) -> 3584 Cores
GP104 (GTX 1080) -> 2560 Cores
GP106 (GTX 1060) -> 1280 Cores
GP107 (GTX 1050 Ti) -> 768 Cores

Wow, yet again, there are 2 dies with more cores and 2 dies with less cores. Middle. If you look at the change in cores when you switch cards, you can choose to go up and gain 50% cores or you can choose to go down and lose 50% cores. That is literally the definition of middle. If you go down again (1060 -> 1050 ti) you lose 40% relative to the 1060 (another 20% relative to the 1080). Which would make 1080 upper-midrange since still 2 more higher SKUs exist above it.

Let's take RAM capacity as another measurement

GP100 (DGX-1) -> 16GB HBM2 RAM
GP102 (Titan XP) -> 12GB GDDR5X
GP104 (GTX 1080) ->8GB GDDR5X
GP106 (GTX 1060) -> 6GB GDDR5X
GP107 (GTX 1050 Ti) -> 4GB GDDR5X

Literally smack in the middle, again. You can even talk memory tech as midrange. 3 techs from a purely bandwidth and capacity standpoint, HBM = fastest, gddr5x = second fastest, gddr5 = third fastest. The middle memory tech.

Memory bandwidth

GP100 (DGX-1) -> 720 GB/s
GP102 (Titan XP) -> 480 GB/s
GP104 (GTX 1080) -> 320 GB/s
GP106 (GTX 1060) -> 192 GB/s
GP107 (GTX 1050 Ti) -> 112 GB/s

Middle. Again.

It's the midrange card, like every x04 card before it. This is beyond nomenclature. The specs are in the middle.

Try it yourself with die size, TFLOPS compute, etc. and see what happens.

I don't want to derail the thread further, but if you got more points please let us hear them

You realize that the cost per mm² is not constant. Right?

16nm is easily 30% more expensive per mm² to manufacture, so that makes your comparisons (28-16nm) not apples to apples.

My point wasn't to compare Pascal (16nm) against Maxwell (28nm), but rather to pinpoint the position of the GTX 1080 in the Pascal product stack.

And by all means, its a mid-end gpu, sporting a high-end price.

I really don't understand why this a problem for some to accept.
And i would have done the same thing if i were the Nvidia CEO, as it obviously worked out great, NV is reporting record breaking profits afterall..

To repeat myself, Nvidia PR/marketing did a splendid job. [Thumbs Up]

 

[railven]

Well the last three or so pages were a waste of forum space.

Logged into say this:

come on guys. There is only one way to look at it. Nvidia is charging these absurd prices and getting away because
1. The consumer is willing to pay.
2. AMD is not providing competition.

Nvidia's gpu revenue increase is staggering. The consumer is paying for those massive profits and AMD is bleeding for its inability to compete.

I don't buy dies, I buy cards. If there is a card at X-Performance for Y-Dollars, when it was AMD making it I bought it. As Raghu's prior post mentioned, GTX 980 Ti upset Fury X at launch. And this generation AMD had nothing to offer me.

Raghu wins this thread. Period.

EDIT: Oh more on page 5 went up while I was posting.

Demand was obviously extreme when taking into account the early supply shortages despite the significant price gouging early on by retailers which Nvidia benefitted nothing from. Based on average selling prices Nvidia could have set the MSRP much higher than they did. You think if Nvidia set the msrp for the 1080 at $350 that any retailer would have sold it for that price? They still would have sold for $750+ because that's what the market determined they were worth. Yet on these boards Nvidia is ripping us off and price gouging us because they would rather increase their revenue instead of letting retailers collect all the profits.

Probably why I didn't care that the EVGA GTX 1080 FTW Hybrid cost $730 was because I sold my $700 GTX 1080 FE Voucher for almost $1,000 on launch day.

Supply and Demand is crazy sometimes!

 

[Headfoot]

Yeah, scrolling through the comments people seem to prefer simple arguments like "that GPU is midrange" and "no it is high-end". That, people, is semantics. Welcome in the happy world of linguistics.

Lol. Dude, really? Die size is only one of many measurements that all lead to the same conclusion. It's all already here so im not going to repeat anything.

 

[witeken]

It affects everybody who makes chips for any product. ASP has gone up significantly more than any production cost increases for Nvidia graphics carrds.

If you even look at the top end 2010 iPad SKUs - it was $829 for a 64GB 10" model. Fast forward to 2016,the price has increased for to $1029 for the 10" iPad Pro and $1229 for the 13" model. There is a price increase,but it pales in comparison to graphics cards.

I don't know the exact prices of chips (and a 100-150mm² mobile chip is of course quite another thing; yields go down with die sizie, so double die size means >double the cost + the yields of 28nm vs 16nm are very likely a world of a difference), nor how they affect Apple or Nvidia's pricing strategy (but again the comparison is off because Nvidia sells a chip while Apple sells a complete product), but I just want to make one thing clear.

Using your excuses we should be seeing laptops costing 2 to 4 times more than 6 years ago and consoles too.

I'm not doing any excuses. I just wanted to correct one mistake. But about the processor prices (which indeed stay pretty constant AFAICT): die size has come down considerably (and still, Intel's bad 14nm yields in 2015 had a noticeable impact on their gross margin).

But anyway, it's of course easy to see from Nvidia's 60% margins this quarter (which is a few % higher than they expected, but that's because the higher volume) that they indeed charge premium prices for their products. And Nvidia is a fabless company, so they pay at least twice as much for a wafer as Intel (because TSMC also has to make >50% margins on their wafers, and because TSMC charges per wafer, worse yields than 28nm are felt by them twice as much; while for Intel they only have the raw production costs to pay)!

I find it hard Microsoft would do a shrink which cost them more money.

Remember that power also comes down and if you don't cram in more transistors, you should be pretty fine.

 

[witeken]

I don't want to derail the thread further, but if you got more points please let us hear them

I just checked prices in my country (which are in euros, to be fair), and by no means I would consider the 1080 a midrange product. Not even the 1070; it costs more than I paid for my One m8, but again there's some currency issues here.

And by all means, its a mid-end gpu, sporting a high-end price.

I haven't looked at the benchmarks all that much, to be honest. I read it's about 30% faster than a 980Ti. Well, processor enthusiasts would be pretty happy with such a performance boost .

 

[USER8000]

snip

Yet,you also forget that the enthusiast chips have not increased in price by anywhere as much.

An example:
1.)Core i7 980X - 239MM2. RRP:$999
2.)Core i7 5820K - 356MM2. RRP: $389
3.)Core i7 6800K - 246MM2. RRP@ $434

So the six core chips have plummeted in price.

Now if we want to make Intel look worse,the RRP from maximum enabled chip to maximum enabled chip has gone up from $999 to $1723,ie,Core i7 980X to Core i7 6950X,over 6 and a half years.

But thats the thing - the Pascal Titan X is a second tier salvage product of a 500MM2 GPU which costs $1200. The second tier salvage product of the GTX580 generation which was of a 500MM2 GPU was the GTX570 which cost $330,

So that is a 3.64X increase in RRP. Intel looks rather like a charity in comparison.

Hence,my point still stands,Nvidia has been successful in marketing that price increase,and good for them. What I find is funny is PCMR enthusiasts seem so much in denial how much they are cash-cows now for other areas,just like Intel used enthusiasts as a cash-cow for Atom chips,which were sold at cost or a loss.

Edit to post.

I also think if AMD does get some more competitive chips out there they will do exactly the same,and the same will go with Zen if it works out OK. We saw it with the Fury and the days of the Athlon 64. They will try and monetise enthusiasts as much as they can to go into other areas.

 

[happy medium]

The GTX580 was $500 at launch and used a full 529MM2 GF114 chip and the Titan X Pascal is $1200 and uses a partially disabled 471MM2 GP102 chip . Since the Titan X Pascal does not use a fully enabled chip,it is closer to a GTX570 in some ways,which was $330 in 2010.

That means in reality,we have gone from $330 to $1200.

ok how about this.
The fastest card in 2011 was the $1000 gtx690 and the fastest card in 2016 the $1200 Titan X P.

That means in reality we have gone from $1000 to $1200 in 5 years.

or
the gtx690 with 2 294mm chips, 588mm total, is 200$ less than the not yet released gp100 with a 610mm die. That would make you correct

and

The second fastest card in 2011 the gtx680 with a 294mm chip was $500. Now in 2016 the second fastest card the gtx1080 with a 314mm can be had for $600.

That means in reality we have gone from 500$ to $600 in 5 years.

 

[nurturedhate]

Die size, power consumption does not matter man.

price/ performance is what dictates what tier a card is in.

In your world a 800mm chip , with performance of a gtx960 is high end.
That's absurd!

Not even close. We already have 960 performance in much smaller packages so obviously your "800mm 960 perf" comment is simple trolling. It all comes down to relative performance within the process node. By your logic Nvidia could start only releasing ever smaller dies at 5% performance increases every two years and that would be the new logical high end even though the ability to produce faster cards is completely possible.

We've been over this repeatedly. When you keep telling everyone the sky is pink with green dots it does not make it true, regardless of how many times you repeat it. 680/980/1080 were never high end. GK110, GM200, P102 all existed when they were released. By any metric 680/680/1080 are 460/560ti successors, same place within the relative stack. You can call it whatever you want, no one cares and it doesn't make you right. Call it upper-mid-high-notaslow-end. Make up words. We don't care. The earth isn't flat. You are wasting time trying to push this stuff.