TechPowerup - Nvidia Kepler GK104 PCB Drawings and power connector pics

[Lonyo]

GK104 is not Nvidia's best card. Sigh...

At the moment, it's the only card, and it's not even out yet...

Fairly sure that makes it the best.
And the worst as well.

 

[boxleitnerb]

You know what I mean. Don't act as if you don't.

 

[railven]

Just trying to get my head around how you think cards are ranked, no need to get immature with the whole "kettle" thing.

But that's my pet name for you

No reason why you need to understand why I see it that way, it has no impact on anything, again, outside of just how I see the cards.

Chillax, kettle

Technically we are paying the same or even less, the money is simply worth less.

At least the money controlled by governments.

If you use non government regulated money, like gold, 1 ounce gold coin that could buy you a $400 card 5 years ago can now buy you 3x 7970.

Don't have to tell me that twice. With how much I spend already in just normal day stuff, I'm suppose to believe my dollar has more value in the PC world? People need to step out of these forums and see what's going on around us. I'm be grateful videocards aren't going up in price the same way other products us.

I'm still only paying $50 more than what I paid in 2010.

 

[Lonyo]

You know what I mean. Don't act as if you don't.

You mean that you expect NV will probably release a higher spec GPU before the end of 2012.

 

[OILFIELDTRASH]

I have a hard time believing 40-50% better performance than gtx 580. Never know though. Maybe this is what we should be getting and 7970 really is bulldozer part 2? I guess only time will tell.

 

[railven]

I have a hard time believing 40-50% better performance than gtx 580. Never know though. Maybe this is what we should be getting and 7970 really is bulldozer part 2? I guess only time will tell.

So then by your logic, it will be $600! Guess you can't afford that one either, eh

 

[boxleitnerb]

You mean that you expect NV will probably release a higher spec GPU before the end of 2012.

Yes. In fact way before the end of 2012. I guess it will be like 460/480 just that they switched places this time. 460 came about 3 months after the 480.

 

[tviceman]

I have a hard time believing 40-50% better performance than gtx 580. Never know though. Maybe this is what we should be getting and 7970 really is bulldozer part 2? I guess only time will tell.

Agreed, probably half that in most situations at playable frame rates. So 20-25% faster than gtx580. Lets see if Nvidia price gouges the crap out of this product, or if they price it somewhat in-line with past mid-range chips.

 

[RussianSensation]

Okay, so using your scenario - if GTX 460 is launched as GTX 480 - is it mid range? All specs are equal minus the name?

Exactly. Had NV launched GTX460 back then and called it GTX480 and called GTX480 == GTX580, GTX460 ("GTX480 as you want to call it hypothetically) is still mid-range. The only thing that matters is HOW much GTX460 costs and HOW fast it is relative to the fastest card in that chain. You are fixating way too much on the name of the SKU or its launch date.

Using your logic, since HD7770 launched at $159, it's AMD's new mid-range card at the moment since the next level up is a $460 HD7950. If HD7970 was launched 3 months later after HD7770 launched, you'd say that's AMD's new high-end HD7000 series card (until HD7970 is released)?

GTX460 is a mid-sized GPU, with performance roughly equal or better than previous NV high-end GTX280/285. That's in-line with NV's strategy of mid-range GPU. GTX460 is by all definitions a mid-range product and its pricing reflected that as well.

And you are again speaking in hindsight. If during the GTX 4 series, what we now know as the GTX 460 - was the GTX 480, would you still call it mid range in retrospect to nVidia's offering?

Not speaking in hindsight. Until you realize that NV has delivered a new upper-mid-range card with performance similar to previous high-end card, this discussion is going nowhere. I don't care if they call it GTX640, 670Ti or GTX880. Price and performance relative to other cards within that generation from NV will determine its own standing. If NV released a new upper-mid-range chip that beats GTX580 (and that automatically means HD7950), that's just what's expected in the first place based on NV's performance improvements going back GeForce 3. NV developed Kepler for 2-3 years. It's not their fault HD7950 is barely faster than Fermi. They probably expected HD7950 to be at least 30% faster than GTX580. Whose fault is that?

Normally when a card beats another, the price is reduced - normally and with age. Not sure WHY GTX 580 saw no price drops, but hey - the markets sustained it. Doesn't mean HD 7970 will maintain its price point.

Why would it? It provides similar performance to HD7950. HD7950s cost $460. You can find GTX580 for $430. 5% performance delta. NV is playing the same game AMD is. If consumers are willing to pay, we will price it as such. I think GTX580 and HD7950 are both overpriced because by now we should have had 30-50% more performance over GTX580 @ $500. How is HD7950 doing on that front? Oh right.

Define generation?

Either a significantly improved previous design (i.e., 4870 --> 5870) or a brand new architecture, in either case accompanied by a performance increase of at least 40-50% above the previous best high-end card. This has been true going back to Radeon 8500 at least. The only exceptions are rebrands such as 9800 series or the flop that was HD2900XT.

Again, until that happens (ie GK100 launches) would you still call GK104, if it launches with the GTX 680 name, a mid-range part?

If GK104 is NV's fastest card that generation (regardless what AMD brings), that's NV's high-end chip.

If GK104 is 50% faster than HD7970 but NV has GK110 that's 50% faster than GK104, GK104 is a mid-range NV chip.

It's not how it does relative to AMD, but how it does relative to its own generation. If GK104 beats HD7950, is it a high-end 28nm Kepler chip? It can still be a mid-range chip in NV's stack if GTX690 blows it by another 30-50%. All it means is AMD didn't bring enough to the table, so they have no business calling HD7950 a high-end chip within the context of a 28nm generation. And therefore, it has no business being priced at $450 since it's only giving us upper-mid-range level of performance based on what's expected from a brand new architecture + node shrink.

Don't blame me for moving the goal posts that define a new generation. I am not moving anything. It has been that way for NV for 10 years. Go blame AMD for releasing a next generation line-up that barely beats last gen 40nm cards without 30-40% overclocks.

At the end, all that matters is GK104's performance and price.

 

[GaiaHunter]

So basically you are saying AMD only makes mid range cards.

 

[RussianSensation]

So basically you are saying AMD only makes mid range cards.

makes is simple tense. Simple tense is an "eternal" state, defined by regularity and frequency of the event occurring. Has AMD released only mid-range cards in the past? No. How did you even get that out of my posts? What you meant to ask me if I think HD7950's level of performance is only good enough to be called next generation upper-midrange? For me, Yes (Post #20). Fast forward 6 months ago before we knew how fast HD7900 series would be. If you asked me where I thought HD7870 and GTX660Ti would end up, I would say ~ GTX580 level of performance. In much the same way HD6870 displaced HD5870, HD4850 displaced HD3870, X800XL displaced 9800XT, etc.

HD7770's performance and that card requiring an overclock just to match a $120 HD6850 already is a sure sign that maybe just maybe the entire HD7000 generation is priced this way only because it has first mover advantage. Another way to look at it, AMD will milk the market while it can, but once NV launches, the market will revert back to the expected historical context of Moore's law vs. performance. Otherwise we are in the hi-fi headphone land of faster cards costing more and more over time.

I am saying for the first time since HD2900XT series (where NV's mid-range 8800GT/S beat AMD's high-end), AMD released what normally would be considered "next generation upper-midrange (HD7950) performance" and priced it as if it's a high-end GPU of that generation. Now, since people are buying them, I can't blame AMD. All it shows is how smart AMD is trying to push $500 GPU with upper-midrange performance on consumers that think it's OK. On the surface, that is not at all how the GPU market has worked in the last 10 years.

Should NV have priced 8800GT at $499 since it blew the doors off 2900XT (AMD's high-end)? It was NV's mid-range part.

 

[Vesku]

Because in all metrics except for Physx and CUDA the 7950 is the better buy. What we are seeing with the GTX 580 pricing is the effect of NVIDIA branding combined with no apparent drive to clear out inventory. What I would expect to necessitate a 10-15% price gap is in actuality a 5% price gap, which makes perfect sense from a strong brand perspective.

Also in terms of branding, I do think NVIDIA is under some pressure to deliver more performance/$ when it comes to the 28nm/Kepler release. Perhaps what we will see is a 670ti at ~$400-450 that fluctuates in performance from 7950 to 7970+ levels depending on game title.

Why would it? It provides similar performance to HD7950. HD7950s cost $460. You can find GTX580 for $430. 5% performance delta. NV is playing the same game AMD is. If consumers are willing to pay, we will price it as such. I think GTX580 and HD7950 are both overpriced because by now we should have had 30-50% more performance over GTX580 @ $500. How is HD7950 doing on that front? Oh right.

 

[Grooveriding]

At the end, all that matters is GK104's performance and price.

Think it is just the latter at this point. We know roughly the die size, bus, power threshold and memory allotment. Card somewhere from just below a 580 worst case, to 7950 levels.

I think at this point all that will matter is price to most people. There won't be any new performance levels being brought to the table, so what it costs will distinguish whether the card is considered 'good' or not, so long as you're basing your opinion on how it's priced compared to what is out there with the same performance and its price.

Also, I'm too lazy to dig up too much - but the 460/285 were close to on par if you take the wins for one or the other and the ties. http://www.anandtech.com/bench/Product/317?vs=313&i=188.191.192.194.195.200.201.216.217.219.220

If NV does still have poor perf/mm² compared to AMD on 28nm, then I can see the card being $300 and performing as a 580. If they have improved their perf/mm² though and brought it up to the level AMD is on in that area, making the card competitive with a 7970, then the card is going to wind up costing significantly more and come in above the $400 price point. I guess it will be price/performance as you said, that determines not just how you might choose to classify the card, but how much it ends up costing.

It's the clock speeds that are going to do it. The 7970, being conservative, is near universally sitting on a 20% performance gain with a slide of the bar in CCC. I would expect nv to be able to hit similar clocks on the same process, but have the advantage of knowing where the bar is already set. What do you do ? Release a card and market it like the 460 was as a great overclocker for a lower price, or put it out 200mhz faster, competitive in benches with AMD and with $100 higher price tag ? D:

Maybe they don't care and won't try to market against the competition and just put out a $300 GTX 580 and say wait till later this year for the big one. Those tri/quad-sli connectors, 670ti rumours and the value of looking competitive in the high end really make me doubt it though.

 

[GaiaHunter]

Only makes implies simple tense. Simple tense is an eternal state, defined by regularity and frequency of events. Has AMD released only mid-range cards in the past? No. How did you even get that out of my posts?

I am saying for the first time since HD2900XT series (where NV's mid-range 8800GT/S beat AMD's high-end), AMD released what normally would be considered "next generation upper-midrange (HD7950) performance" and priced it as if it's a high-end GPU of that generation. Now, since people are buying them, AMD is loving it. But on the surface, that is not at all how the GPU market has worked in the last 10 years.

That's like saying, NV should have priced 8800GT at $499 since it blew the doors off 2900XT (AMD's high-end).

Because generally AMD high-end cards are the size of the NVIDIA mid range cards?

They are just generally faster and have more features than similar size NVIDIA cards and so are able to compete with NVIDIA high-end cards, that are much larger.

Occasionally, these mid.range cards are actually out of the doors much sooner than NVIDIA cards, and so they are able to take a lead over the previous NVIDIA high-end cards and charge high end prices, see the 5870 and 7970.

 

[RussianSensation]

Because in all metrics except for Physx and CUDA the 7950 is the better buy. What we are seeing with the GTX 580 pricing is the effect of NVIDIA branding combined with no apparent drive to clear out inventory. What I would expect to necessitate a 10-15% price gap is in actuality a 5% price gap, which makes perfect sense from a strong brand perspective.

I am not in any way saying that GTX580 is a better buy. I think its price should have fallen to $380 last year already. I'd buy HD7950 in a heartbeat over the 580 @ $450. But the only reason HD7950's price even makes any sense is when put into context with the 580. The problem is the 580 is a 15 months old card, on 40nm tech, using 2 year old architecture. Is it reasonable to expect faster performance @ similar price or much lower price @ similar performance? I think so. Why? Because it has happened for as long as anyone can remember. Using the 580's price to justify that HD7950 is reasonably priced undermines the very idea that the 580 itself is simply way overpriced for the level of performance given the time since its own release.

How can you have $220-250 GTX480 on Newegg and GTX580/HD7950 for $440-460? 2x the price increase for 20% more performance from a new generation? You don't see anything odd about that?

But if the market participants are willing to drop $450-500 on GTX580 level of performance 15 months later, I say AMD has every right to take advantage of those consumers.

 

[AtenRa]

Because generally AMD high-end cards are the size of the NVIDIA mid range cards?

They are just generally faster and have more features than similar size NVIDIA cards and so are able to compete with NVIDIA high-end cards, that are much larger.

Occasionally, these mid.range cards are actually out of the doors much sooner than NVIDIA cards, and so they are able to take a lead over the previous NVIDIA high-end cards and charge high end prices, see the 5870 and 7970.

With GCN and DX-11 games that will not be the case this time. Tahiti will not be smaller and faster like HD5870 was.

 

[RussianSensation]

Because generally AMD high-end cards are the size of the NVIDIA mid range cards?

What does the size of AMD vs. NV have to do with anything? It's been explained that GK104's size is important relative to NV's own strategy. We have no idea how GK104 performs against a 350mm^2 Tahiti XT. How does that preclude NV from launching a 500+mm^2 high-end card? Why do you have to bring AMD vs. NV die sizes into this when the discussion is about NV's GK104 die size vs. how that fits into their own line?

They are just generally faster and have more features than similar size NVIDIA cards and so are able to compete with NVIDIA high-end cards, that are much larger.

Consumers don't care about die sizes, unless die sizes negatively affect the end product such as having negative consequences that may include high power consumption, or delayed product due to lower yields as a result of a large die size, etc. Once the product is on the shelf, they look at the price, if their PSU can handle it, how much performance they need, what features they want, etc. No one goes up and asks: "Excuse me sir, does this HD7970 have a smaller die size than a GK104 or not?"

Regardless, your die comparison should be thrown out the window since NV has been dominating discrete GPU market share since GeForce 8. Just more evidence that most consumers don't care about die sizes in the end product.

GeForce 3 Ti 500 > Radeon 8500
GeForce 4 had 0 competition from ATI
9700Pro/9800XT > GeForce 5
GeForce 6 and X800XT traded blows (AMD winning in shader intensive games, NV in Open GL)
GeForce 7 and X1800/X1900 traded blows (AMD winning in shader intensive games, NV in Open GL)
GeForce 8 blew the doors off HD2900 and 3800 series
GeForce 200 traded blows with HD4800 (GTX280/285 > 4870/4890)
GeForce 400 traded blows with HD5800 (It took HD6970 just to match the performance of GTX480)

To state that AMD generally has faster cards is 100% incorrect. AMD lost the single-GPU performance crown for the last 3 generations in a row.

To state that AMD generally has faster cards and to slide in the phrase "per die size" is like saying Porsche has faster cards than Ferrari per cylinder. No one compares how awesome Ferrari or Porsche is "per cylinder". When people buy a Ferrari they don't go, oh it has a larger engine, nah I'll get the Porsche instead since it has a smaller engine.

You think if 500mm^2 HD7980 was 50% faster than HD7970 people would stop buying it because "Oh no, it's a large die!! I hate large dies".

 

[96Firebird]

They are just generally faster and have more features than similar size NVIDIA cards and so are able to compete with NVIDIA high-end cards, that are much larger.

What features are these? Going forward with 28nm GPUs, Kyle at [H] said Nvidia now has Surround with a single GPU. I'm willing to believe him on this, since they should have done this a while back.

 

[Crap Daddy]

What features are these? Going forward with 28nm GPUs, Kyle at [H] said Nvidia now has Surround with a single GPU. I'm willing to believe him on this, since they should have done this a while back.

Kyle also said this is 45-50% above GTX580 in artificial 3D benchmarks. Was this mentioned here before?

 

[RussianSensation]

Maybe they don't care and won't try to market against the competition and just put out a $300 GTX 580 and say wait till later this year for the big one. Those tri/quad-sli connectors, 670ti rumours and the value of looking competitive in the high end really make me doubt it though.

But if GTX670Ti is really their high-end card, how can a 350mm^2 chip with 256-bit memory bandwidth beat a GTX580 by 40-50%? Unless they are using some PhysX benchmarks or Unigine Heaven with Extreme Tessellation to prove their point.

A GTX670Ti at 150% over GTX580 is like 20% faster than an HD7970. That sounds too good to be true considering AMD has a track record of making very fast chips on a smaller die. My guess is GTX670Ti will cost less than HD7970 and come in between HD7950 and HD7970 in performance. Ultimately, overclocked vs. overclocked HD7970 will have a lead due to its massive memory bandwidth advantage.

 

[Grooveriding]

But if GTX670Ti is really their high-end card, how can a 350mm^2 chip with 256-bit memory bandwidth beat a GTX580 by 40-50%? Unless they are using some PhysX benchmarks or Unigine Heaven with Extreme Tessellation to prove their point.

A GTX670Ti at 150% over GTX580 is like 20% faster than an HD7970. That sounds too good to be true considering AMD has a track record of making very fast chips on a smaller die. My guess is GTX670Ti will cost less than HD7970 and come in between HD7950 and HD7970 in performance. Ultimately, overclocked vs. overclocked HD7970 will have a lead due to its massive memory bandwidth advantage.

That's taking it too far though. The GK104 is most definitely not their absolute top card. A 670ti leaves the 680 name wide open and if it performs 25% better than a 580, would qualify as fitting in to the 670 bracket. You can't have it both ways here, is the GK104 a GTX 580 performing card with a 350mm² die for $300 or is it a GTX580+20-25% performing card with a 350mm² die for $450 ? Or do you think they are going to break with their history that you have been using to make points on and release a card that is GTX580+20-25% for $300 ? Personally, I think it is the happy_medium right in the middle of those three scenarios.

I expect the GK110 to be the card that is 150% of a 580, and called a 680 or whatever. A 670 being 20% or 25% or even 30% faster than a 580 is just fine. 50% is about what you see on average with 480 over 285 with some outliers above or below that mark, I don't expect we'll see much different now on 28nm.

Of course I do think they will break GK110 into more than one card. Unless the plan is GK104 is the 670 and GK110 is the 680. The only guess I will give with certainty is that nvidia will not release a card as fast as a 7970 for $300, ever. If we are going to keep using nvidia's history, I don't remember the grounds for a mid-range card as a first release at a cheap price point.

 

[Vesku]

Is this not a given? Again going to place my bet on a ~$450 card that floats between 7950 and 7970. It may not even force AMD to readjust prices, as it could once again run a bit hotter and need a heftier cooler.

I agree though, with 40nm->28nm we are seeing a change from the trend. The performance/$ gain with new nodes appears to be slipping.

Unless they are using some PhysX benchmarks or Unigine Heaven with Extreme Tessellation to prove their point.

 

[GaiaHunter]

Die size explains performance.

If a card is bigger die size and slower/similar speed, it is a turd.

IF a card is bigger die size and faster, it is explained by its bigger die size.

If a card is faster and smaller its great or the competition is a turd.

If NVIDIA K104 is similar die size and as fast as 7970/7950 it is only natural.

 

[RussianSensation]

Die size explains performance.

I didn't know that. Maybe next time instead of checking specs of GPUs, I should just make a list of their die sizes and buy based on that.

If you can tell how fast a card is based on its die size, how fast is GTX670Ti? You can guess since you know its die size is about 350mm^2. How fast is its Tessellation performance over HD7970? How much better is its performance in games with deferred MSAA? Can it handle 2560x1600 resolution better than GTX580 did?

Die size by itself tells you nothing about performance, power consumption, or the price of the card. At best, it's a guestimate to where those things might end up based on performance of previous generations. Even then, node shrinks allow more transistors to fit into a given die size and different 28nm node manufacturers can fit different amounts of transistors. Different architectural designs can also dictate how many transistors can fit even in the same node.

If a card is bigger die size and slower/similar speed, it is a turd.

So if GTX580 cost $350 right now, it would be a turd right now vs. the HD7970? Price, features, etc. those things don't come into your purchasing decision, but die sizes do? Interesting.

IF a card is bigger die size and faster, it is explained by its bigger die size.

Not necessarily. There are too many variables. What if the the larger die card is clocked 20% faster? Did it win because of its larger die or because it was clocked 20% higher? Or because its architecture was more efficient? What if 1 card has extra features used in professional markets that are not used for games? The die size would need to include those features even if they are not used in gaming markets. Your world is too black or white. I would say most consumers don't care about die sizes. I would be most of them look at the price first. Enthusiasts who don't care about the price look at overall performance and overclocking abilities. AMD's market share isn't showing that consumers are attached to the idea that die sizes are important. Maybe because die sizes is probably the 10th item on the list for people who are in the market for a new GPU?

So in your world a 5.0 Litre muscle car is a turd because it didn't go with a 3.5 Litre turbocharged engine instead to make the same horsepower? Why is it wrong that 2 companies use entirely different approaches and that one has to be better than the other? And if AMD's approach is better, neither its market share nor financial results are showing that it is better either. So really, what's your argument about die sizes, that you care? That's fine, just don't say the rest of the market does. But the rest of the market sure does care about performance for the price, regardless of die sizes.

If a card is faster and smaller its great or the competition is a turd.

So things like price, features, game bundles, warranty, overclocking abilities, performance in games you personally play don't matter over the more "important" die size?

By your definition NV hasn't made a good card since.... I don't know.... for a long time. I guess GeForce 4200 Ti, 6600GT, 6800GT, 7800GT, 7950GT, 8800GT/S, GTX460, GTX560/560Ti/570 are all crappy?

If NVIDIA K104 is similar die size and as fast as 7970/7950 it is only natural.

Let's say hypothetically, GK104 is 500mm^2 (it's probably not though). If it's as fast as an HD7950 and costs $399, it's a turd?

I love how in your world price and features are never a consideration, but die sizes are key for extrapolating performance. And yet you can't even tell how fast GTX670Ti is based on its die size.

 

[exar333]

Die size explains performance.

If a card is bigger die size and slower/similar speed, it is a turd.

IF a card is bigger die size and faster, it is explained by its bigger die size.

If a card is faster and smaller its great or the competition is a turd.

If NVIDIA K104 is similar die size and as fast as 7970/7950 it is only natural.

This is one of the most unintelligent posts I have ever read on AT.