TechPowerup - Nvidia Kepler GK104 PCB Drawings and power connector pics

[railven]

Its about the same size.

Looks like March 23rd launch.. which means we will see some performance leaks very soon.

I'm giddy with excitement. I'd wish for a sooner release date, she can hold off until then.

I however am buying the moment the card I want is in stock at the egg since I got some "coups" to burn.

 

[GaiaHunter]

I thought these leaks showed a die that was smaller than Tahiti.

People are saying around 340 mm² from eye balling vs the 360 mm² from Tahiti.

Or did I miss some rumour?

 

[tviceman]

I thought these leaks showed a die that was smaller than Tahiti.

Its about the same size.

People are saying around 340 mm² from eye balling vs the 360 mm² from Tahiti.

Or did I miss some rumour?

No, you got it right. Tahiti is ever just so slightly larger. Also, Tahiti also has more of it's die dedicated to it's larger memory controller.

 

[f1sherman]

Ok, now I'm getting excited. If I can get 7970 performance (or better) for 300, I'm not even going to hesitate to go over to team green this round.

reportedly 45-50% faster then 580 in Artificial 3D benchmarks

so I wouldn't get excited just yet

 

[Homeles]

After AMD acquired ATI and decided to focus on the small-die strategy:

2900XT (huge flop)

The 2900XT was probably the reason why AMD went with the small die strategy... it was released prior to the strategy change.

 

[TakeNoPrisoners]

reportedly 45-50% faster then 580 in Artificial 3D benchmarks

so I wouldn't get excited just yet

I don't judge performance until I see a wide variety of game results. A couple game benchmarks or artificial benchmarks don't really mean much.

 

[MrK6]

I don't judge performance until I see a wide variety of game results. A couple game benchmarks or artificial benchmarks don't really mean much.

This. Actually, I don't even bother looking at canned benchmarks anymore either. There's enough optimization that they're not even close to what real-world results are. I wish more sites would take [H]'s approach to benchmarking, as they're always the closest to the experience I actually get.

 

[Silverforce11]

reportedly 45-50% faster then 580 in Artificial 3D benchmarks

so I wouldn't get excited just yet

No, the 680 is reportedly 50% faster than 580 in canned benchmarks.

We don't know if thats gk104 or gk110.

 

[Sind]

People keep saying reportedly, who are they and where are they reported?

Edit oic Kyle, with no links. Something concrete would be nice.

 

[RussianSensation]

It is a smart move on nvidia's part being late with their flagship. They clearly learned and came up with a better strategy after going through the same issues on 40nm. Had the 460 come out in Jan/2010 and the 480 in a better state in August/2010, there would not be the blight of JHH and his grilling machine, wood screws and all.

True. GK104 might even be a $249-299 mid-range part. For 6 months now we have heard rumors that in Feb-April NV plans to release GK104 (which by all previous rumors was not NV's high-end card). It was meant to be a replacement for GTX560 Ti. I have no idea why so many people are assuming that GK104 is NV's high-end for Kepler generation. It makes a lot of sense if GK104 is a mid-range part and launches to compete against the market where the most sales are taking place --> against HD7850/7870.

The fact that the leaked GK104 engineering sample has 8 memory chips likely suggests a 256-bit card. It also has a 6+6 pin connector and not a 6+8 pin connector designated for high-end GPUs. How can a 256-bit card with 6+6 pin be High-end? It doesn't make a lot of sense.

It should be "LOL it took them x months and a new node for AMD's 350mm2 part to beat NVIDIA's 500mm2 part.

I guess it comes down to is GK104 @ 340-350mm^2 the new high-end GPU or will there be another 500mm^2+ high-end GK110? If GK104 is just a GTX660/670Ti style card, then the large die strategy continues and most surely if that happens then AMD might lose the high-end single GPU crown.

>> In relation to this thread, I am really excited for when the GK104 arrives. Because it will be the best showdown between AMD and NVIDIA in a long while. In terms of pricing, power draw, cooling solution, basically everything. This is not the "big as it will" go NVIDIA product, that will deserve it's own discussion when there is more substantial information about it.

Agreed.

If bigger die is soo much better, why the sudden turn from nvidia?

Well we still don't know if GK104 350mm^2 is the highest end card though. It could be. Then it would signal a turn of strategy at NV.

 

[RussianSensation]

AMD acquired ATi in 2006, you think AMD really had much to do with HD2900 which was released in 2007? I'm guessing HD2900 was pretty much set in stone and getting ready for manufacturing shortly after the acquisition. It also wasn't a small GPU, it was a pretty big one.

You are right.:thumbsup: Thanks for correcting me. I just looked up the die size of R520 = 420mm^2. That card also had a 512-bit memory bandwidth bus. Looks like they miscalculated a lot of things with that design.

My point is if the small die strategy has worked, why has AMD's GPU division lost $ for so many years, relinquished discrete GPU market share to the current 63%, possibly eroded the brand value of AMD to a "price/performance brand", etc.?

NV could have chosen to pursue the small die strategy as well. Their engineers come from Waterloo, MIT, Stanford, etc. NV probably didn't do that because it didn't make any sense for them at the time. And 5 years later since the large die G80, they have made more $, gotten more market share and won the single GPU crown in every consecutive generation. Not sure how the small die strategy is working for AMD on the desktop market. It's certainly a better fit for mobile/APU strategy.

you must be forget one tiny detail, the small die strategy let AMD regain market share, and do you forget about HD 5970 ??? that card is the fastest card for like a year and a half. and then HD 6990 is toe to toe with GTX 590.

AMD lost market share from about 50/50 to 36/63. So I am not sure how the small die strategy allowed AMD to gain market share.

HD5970 was a great card. HD6990 was a very loud card (2nd loudest card of all time IIRC). But not sure how those cards matter in the single-GPU context.

Geez RS. I usually find your posts well thought out and interesting but AMD really seems to have gotten under your skin with the 7900 series. I (like everyone else) wish the cards had released at lower prices too but you come off like there is no value in them. I gamed on a pair of 6950's for a year and to be able to go to a single card that performed the same or better is great in my book.

I also recommended HD4890 over GTX285 and I recommended HD4870 over GTX280 as well as HD6950/GTX570 over GTX580.

The reason HD7900 series has disappointed me so much is because AMD is now neither delivering its awesome price/performance of HD4800, 5800 or 6900 series, neither the 40-50% faster performance over the previous generation high-end card (which is GTX580 @ $500). To me it makes no sense to say that HD7970 brings 40% over HD6970 because it's not a $379 card.

So if it's a $500 card, I expected it to be 40-50% faster than the previous "generation" $500 high-end card == that is GTX580.

That's achievable with overclocking to 1200mhz. I personally think it should have been 40-50% faster "out of the box".

Basically, if Kepler's $500-550 high-end card is not faster than the previous generation $500 high-end card (i.e., that means GTX580 from 40nm), I'll also be disappointed.

I don't view it as the performance increase of AMD to AMD or NV to NV. It's about generation leap vs. the previous best card from either brand in the last generation. The previous high-end card was GTX580 and HD7970 is just 25% faster than that without an overclock.

But anyway, back to GK104. As you said all we know is die size and 256-bit memory bus at this point. Doesn't tell us much at this point.

 

[Silverforce11]

Not official but all signs point to gk104 being a mid-range chip thats to take on the role of high-end product simply due to delays of the real big Kepler.

They can clock it higher than designed specs to make it fill that role better. We will know from the OC headroom.

 

[RussianSensation]

First you forget this are technical forums and so it is possible to talk about engineering which you seem to have problems with, but you have no problems with economics, as if all this isn't related.

I have no problems discussing die sizes. It's you who said a card with a larger die size and similar speed is a turd. I provided GTX460 and 8800GT as examples of excellent cards. Larger die sizes than AMD's 6850 and 3870 and far more successful videocards.

By the way, AMD did make a slightly larger 3870, called the 4870/4850 that beat the 8800GT senseless

Are you serious? To prove your die size point you use a 55nm next generation architecture HD4870 (there were major redesigns for AA shaders among other things in that architecture over 3870) to compare to a mid-range G80 65nm 8800GT to show how AMD is able to make a better card at a smaller die?

Your key point that AMD's engineers are "better" because decided to make a smaller GPU is odd to be honest. You make it sound as if NV's engineers don't have that ability or failed to consider that NV deliberately decided NOT to make small die GPUs because it makes no sense for them.

If NV makes a 500mm^2 GPU that has better price or performance or price/performance than a 300mm^2 AMD card, that's what that matters and what sells cards. Yes, sure, AMD's engineers made a more efficient GPU. So what? If it doesn't sell better in the market, makes more $ for AMD, gains more market share, it's ONLY a check-mark achievement for engineers. The engineers should pat themselves on the back and next time make a larger die card to actually beat NV because outside of the APU market, not many people care that they made a more efficient GPU that ultimate lost. It would only matter if they won (financially, market share, brand recognition, etc.).

If right now AMD released a 550mm^2 HD7980 with 40% more performance than HD7970, you'd call that a failed strategy?

The point of the matter is consumers compare cards based on price, features, image quality, performance, not die sizes.

If you want to discuss the merits of engineering, AMD currently makes more efficient GPUs per mm^2. That's not something to argue about -- that's a fact until NV releases something more efficient. But in any given generation, NV always has similar cards to match AMD in terms of price and performance. So essentially, while AMD has more efficiently engineered GPUs, it means squat in the end. Sure, their engineers have thus far made more efficient desktop GPUs. As I said they have to because of the APU strategy. So what? What has their "superior" knowledge of making more efficient GPUs done for AMD on the desktop market? Not much.

Is a 6.3 Litre supercar (Ferrari F12) worse than a 3.8 litre twin-turbo supercar (McLaren MP4-12C)? They achieve similar performance but use a different approach to engineering. In your world the more efficient one is always better. Black or white?

You make it sound as if a small die GPU is some advantage. In what way exactly? To say that AMD's engineers are "better"? That's your main point? Did it ever occur to you that NV and AMD engineers leave and work at each other's firms and that the GPU design direction is simply different? Why would NV (Ferrari) decide to stop making 6 Litre engines when it worked well for them all these years? Why would AMD (McLaren) decide to start making large engines when making small engines worked well for them?

See how the engineering design choice does not in any way imply 1 is better. They are just different. Small vs. Large die strategies are just different. But so far NV has executed it much better, made way more $ since G80, has 63% market share in discrete GPU segment, which leads credence to the fact that small die strategy hasn't lived up to the hype on the desktop.

NV could have easily done the same in 2007, but obviously they went with a large die strategy at the time because it made sense for them to focus on what they are good at = making massive powerful GPUs to service GPGPU, gaming and professional markets.

 

[GaiaHunter]

I have no problems discussing die sizes. It's you who said a card with a larger die size and similar speed is a turd. I provided GTX460 and 8800GT as examples of excellent cards. You countered 8800GT with a next generation HD4870 made on a smaller node?

Excellent cards cause of the price and mostly timing.

You said the 3870 while smaller didn't get a pass vs the 8800GT and if AMD couldn't make a bigger card that was their problem. The 3870 and the 4870 were on 55 nm and the 8800GT was on 65 nm.

If NV makes a 400mm^2 GPU that has better price or performance or price/performance than a 300mm^2 AMD card, that's all that matters. Yes, sure, AMD's engineers made a more efficient GPU. So what? If it doesn't sell better in the market, makes more $ for AMD, gains more market share, it's a check-mark achievement for engineers. The engineers should pat themselves on the back and next time make a larger die card to actually beat NV because outside of the APU market, no one cares they made a more efficient GPU that ultimate lost. It would only matter if they won (financially, market share, brand recognition, etc.).

So are you implying those are factors when you deciding which card is best to buy?

I see a 6950 on your sig.

And did you notice the most successful NVIDIA card last gen was the GTX460? Not a >400 mm^2 is it?

Timing is very important.

The point of the matter is consumers compare cards based on price, features, image quality, performance, not die sizes.

Really?

How many times are you going to repeat this when this isn't even at discussion.

Where did I say people buy using die sizes?
Don't bother, cause I didn't.

You make it sound as if a small die GPU is some advantage. In what way exactly? To say that AMD's engineers are "better"? That's your main point? Did it ever occur to you that NV and AMD engineers leave and work at each other's firms and that the GPU design direction is simply different? Why would NV (Ferrari) decide to stop making 6 Litre engines when it worked well for them all these years? Why would AMD (McLaren) decide to start making large engines when making small engines worked well for them?

You not understanding.

If I can pack more performance in a smaller die it is better.

No doubt about that.

That has nothing to do with AMD strategy.

The reason AMD can compete with NVIDIA is due to its higher performance/mm^2 since they aren't going for bigger dies.

Why isn't AMD building bigger dies? No idea. Might be related to the fact they are going for APUs, 30 million of them made on 40 nm.

Maybe they are counting on it being faster to develop and reach a point where the gaps that already exist since AMD started to focus on smaller dies will keep widening (on their opinion).

See how the engineering design choice does not in any way imply 1 is better. They are just different. Small vs. Large die strategies are just different. But so far NV has executed it much better, which leads credence to the fact that small die strategy hasn't lived up to the hype.

I could see if I actually said small dies are superior to big dies. I didn't said that. I said higher performance per mm^2 is better, in every aspect, which you concur.

Which is somewhat similar to higher IPC is better than more cores.

It isn't my fault that you believe I have to be taking sides, although that is common syndrome in these forums.

 

[BallaTheFeared]

The reason AMD can compete with NVIDIA is due to its higher performance/mm^2 since they aren't going for bigger dies.

We'll find out if that's true this gen as this is the first time AMD has really tried to add gpgpu capabilities more in line with what Nvidia has had for awhile.

GK104 while slightly smaller than the 7970, should be a good card to judge this sort of opinion.

 

[GaiaHunter]

It seems to me if GTX670 ti is a single GPU card (and not a GK104x2 that was rumoured by charlie) and is similar performance to the 7970 (and faster than the GTx580 by extension), the only way NVIDIA will price it at $300, and even $400 might be pushing, it is if the release of the bigger dies are imminent as well.

No way NVIDIA will be in the market with no card over $300-400 by their own will (GTX580 would have to drop).

If the rumours about the GK110 being released later in the year are true, it makes no sense. IF NVIDIA purpose is to stall AMD, they might as well have dropped their price on the GTX580 already.

 

[GaiaHunter]

We'll find out if that's true this gen as this is the first time AMD has really tried to add gpgpu capabilities more in line with what Nvidia has had for awhile.

GK104 while slightly smaller than the 7970, should be a good card to judge this sort of opinion.

That is what I've been saying.

 

[tviceman]

I wish more sites would take [H]'s approach to benchmarking, as they're always the closest to the experience I actually get.

That would be pretty terrible actually. Very small selection benchmarks, only 1 apples-to-apples comparison per game at settings which are usually too demanding to be playing at in the first place.

worthless

good

borderline between being smooth and choppy

good

That is all the apples-to-apples from 1 review. That's it. Only 4 games benchmarked. And the only 4 direct comparisons, one is worthless from a usability point of view and another is borderline to many, many gamers.

 

[BallaTheFeared]

30's are too low for me, mid 40's is all I can stand (personal preference).

But yeah, [H] should do SLI/CF reviews at that res or reduce the settings, but how else can you showcase 2gb/3gb of vram without excessive settings on a 1k screen res?

 

[RussianSensation]

Gaia, I think if you want to discuss die sizes, let's take it to PM because it's not contributing to the specs of GK104.

 

[Silverforce11]

2560 res gameplay for top end cards... Why is that worthless?

If its running 40 fps at that res and max details, for someone with 1080p, it will fly. If you want to see every res there's plenty of places where its done. [H] is specifically targeted towards high-end and his fps graphs are more meaningful than every other site that just does min/avg/max graphs, which means nothing as jerky fps or fps drop due to loading don't show up in those charts.

He could expand it to include more dx11 games.. but srly, what else, AvP, Metro? They are old and forgotten. Again, plenty of other sites review a plethora of games. [H] sticks to popular and or latest ones.

As to the die sizes, consumers don't care. The only factor stemming from die sizes that may affect some consumers are power use and perf/w, even then, its been a non-issue unless its really terrible.

 

[RussianSensation]

2560 res gameplay for top end cards... Why is that worthless?

I think you missed the main point tviceman was making. Kyle is trying to test "playable" performance but what he really shows is what's "Playable for him" since his standards are so low that often find cards running at 35-42 fps in FPS (BF) or racing games (Dirt 3) as playable. That's not playable in a competitive FPS or racing genre to most people.

Secondly, 4 games is not enough to make a reasonable recommendation if one card is better than the other at the end of the review when other websites bench 10+ games. If he wants to do modern game tests, then he should do reviews on a per-game basis like GPGPU.ru does. His review of HD7770 clearly showed that he arrived at horrendous conclusions due to his often flawed testing methods.

Tom's found that HD7000 series is not very fast for DX9 game engines. They made sure to mention that several times. Would you have known that based on 4 games tested at HardOCP? Not a chance. Not all games are DX11 games. It's important to know that HD7900 series suffers in DX9 engines. Just like BFG found that Fermi wasn't so great for OpenGL game engines vs. GTX285.

For example, if another website only tests 4 games, we use that as a data point but not useful enough to ignore 10 other websites. And yet HardOCP is put on some special pedestal as an exception?

2560x1600 is useful, but testing cards at 30-40 fps or worse at 22 fps makes no sense when your entire review is based on "Playable Settings". That's not "Playable Settings", but Kyle's own playable settings.

H is specifically targeted towards high-end and his fps graphs are more meaningful than every other site that just does min/avg/max graphs

Tons and tons of other websites run custom benches (hardware canucks, gamegpu.ru, pcgameshardware, computerbase.de, bit-tech) and PCPerspective and Drivenheaven even provide similar time graphs to HardOCP. TechReport does custom runs and even provides their own unique testing method.

How come you think HardOCP is better than all of the above when they only test 4-5 games and time and time again their reviews are often way off with with the results achieved by the many testers I listed above? You don't find that odd? You'd rather use the results of 1 website against 8 others that do custom bench runs?

His idea of using timegraphs is great, but then he tests cards all different settings at what he thinks are playable frames and never provides any videos.

Also, he's made very glaring remarks that MLAA > MSAA, that almost all canned benchmarks are useless, when in fact A LOT of canned benchmarks represent performance much in line with real world gameplay (Far Cry 2, Crysis, HAWX 1/2, AvP, etc.). A lot of users here even run canned benches when they upgrade GPUs and no one accuses them of bias.

A lot of canned benchmarks are sections from a game and in fact do represent real world performance. Fine, let him do custom runs, which is even better. But then he tests apples to apples and concludes that one card is unplayable at 35 fps but another is playable at 42 fps. What? If people think it's an improvement to go from 35 to 42 fps in a FPS genre RPG or shooter, they should get a PS3 and be done with PC gaming. So many times he tests cards at 2560x1600 at framerates where often ALL the cards are unplayable for a lot of PC gamers and he calls that playable....while most other reviews are saying that you need 2 cards to actually make those games playable at those settings.

 

[railven]

No.

I am talking about An average user that bought 5850 2 years ago
For like $259.

I am not talking about early adopters,
who are willing to pay no-matter-what

Wasn't that a small fraction of users (most likely first shipment?) I remember I waited on buying HD 5870 launch with the idea that it would drop price when my time to buy came.

It went from MSRP $380 to gouging in the $420s some even $450. Why I jumped on my card when I finally decided and even the HD 5870 2GB went from $480 to $520.

Most users that bought an HD 5k at launch were lucky SoBs that had no reason to upgrade for a very long time.

 

[Grooveriding]

Wasn't that a small fraction of users (most likely first shipment?) I remember I waited on buying HD 5870 launch with the idea that it would drop price when my time to buy came.

It went from MSRP $380 to gouging in the $420s some even $450. Why I jumped on my card when I finally decided and even the HD 5870 2GB went from $480 to $520.

Most users that bought an HD 5k at launch were lucky SoBs that had no reason to upgrade for a very long time.

5870s and 5850s were super hard to get the first few months after the initial launch batch. It was a feeding frenzy with people camping etailors and threads springing up when they were in stock, only to sell out in hours. I would say that is part of why the 7970 is priced more in line with a new high end gpu price.

 

[AnandThenMan]

I think you missed the main point tviceman was making. Kyle is trying to test "playable" performance but what he really shows is what's "Playable for him" since his standards are so low that often find cards running at 35-42 fps in FPS (BF) or racing games (Dirt 3) as playable. That's not playable in a competitive FPS or racing genre to most people.

How do you know what is playable to "most people" are you the ipso facto? Kyle has been doing reviews for about 14 years, how long have you been doing reviews?

I don't always agree with his findings, but one thing I do find is I can reproduce his results fairly closely, plus you can extrapolate from his results what hardware you would need to get what you consider acceptable framerates. Real world gameplay>scripted benches any day of the week.