GeForce Kepler 104 (GK104) Packs 256-bit GDDR5 Memory Bus, 225W TDP

[Vancha]

Going from 170w using 40nm to 225w using 28nm does not necessarily mean monstrous performance. It could be the opposite.

I chose my words very carefully.

 

[punani]

Going from 170w using 40nm to 225w using 28nm does not necessarily mean monstrous performance. It could be the opposite.

Could you clarify ?

Higher wattage on a smaller process is a "two times" performance increase by my understanding.

 

[devguy]

If this is true, then how is nVidia going to pull of another dual GPU card (especially if this isn't even their flagship Kepler)? Super low clocks?

Even if this outperforms the HD 7970, the HD 7990 will walk all over it (and also probably outperform the flagship Kepler if this is not it).

 

[CrAsHnBuRnXp]

zzzzz

Says the fanboy

 

[treboRR]

GF104 to GK104.


Somehow 225W seems all wrong, unless, they are aiming at performance of 4x, which I would say is physically and technically impossible unless their transistor count has gone through the roof, and this chip is the size of a football pitch! (And essentially SLI on-chip).

C'mon man! ATIs 7970 has twice the transistors count (2640M vs 4310M transistors - 6970 vs 7970) but has only 20% or little more perf vs last gen of cards so think twice! ATI failed to sqeeze the perf out of those transistors. So lets hope Nvida will not fail with his architecture

 

[CrAsHnBuRnXp]

Enough talk nVidia just release the damn thing.
People can see for themselves, right now its all speculation.

I read somewhere (might have been here on TPU) that they were releasing the cards in March-April.

 

[Nihilus]

If this is true, then how is nVidia going to pull of another dual GPU card (especially if this isn't even their flagship Kepler)? Super low clocks?.

Yeah performance is great and all, but what will the gtx 680 pull - 300w?! ? And the gtx 690?!!

 

[Steevo]

Could you clarify ?

Higher wattage on a smaller process is a "two times" performance increase by my understanding.

Memory cache is a huge thermal wasteland. If they had to increase cache to gain performance at whatever core/shader speed the accompanying power use has minimal impact on actual performance as they don't scale.

For memory to operate reliably at the same frequencies as the core it has to be balanced between capacitive energy stored and drain, if it stores too much it can't run at high speeds as it won't switch fast enough, and if it drains (leaks) too fast it takes a crap load pf power. So each Kb 1024 bits plus any ECC if needed adds to the overall thermal package, and eats up valuable real estate on the die.

They could have a BullDozer on their hands.

C'mon man! ATIs 7970 has twice the transistors count (2640M vs 4310M transistors - 6970 vs 7970) but has only 20% or little more perf vs last gen of cards so think twice! ATI failed to sqeeze the perf out of those transistors. So lets hope Nvida will not fail with his architecture

Obviously you have ignored the overclocking feats these cards are capable of......plus immature drivers due to a whole new design...


Its only 20% at stock, and once you overclock these monsters they eat the competition and all you have left is a green cloud fart.

 

[Crap Daddy]

Its only 20% at stock, and once you overclock these monsters they eat the competition and all you have left is a green cloud fart.

Interesting way to compare a new arch/generation product on 28nm to a product that's more than one year old tech. Or you want to suggest that this is your opinion about the new unreleased but only talked about NV GPUs?

On another note everything gets more confusing and I really don't think we'll see something worthy before April.

 

[Yellow&Nerdy?]

They have redone the naming scheme I think. The TDP is 225W, which is 15W higher than the 7970: unless Nvidia plans on pulling off another Fermi, that should be a good indication of the performance. Quite interesting that the memory bandwidth is 256-bit, which is off the GF114, not the GF110. Then again I don't blame Nvidia for going further with the GF114 instead of the GF110, because we all remember what a steaming pile GF100 was.

 

[Completely Bonkers]

 

[Steevo]

Interesting way to compare a new arch/generation product on 28nm to a product that's more than one year old tech. Or you want to suggest that this is your opinion about the new unreleased but only talked about NV GPUs?

On another note everything gets more confusing and I really don't think we'll see something worthy before April.

As things go ATI/AMD seemingly looked at what the competition was doing and made their own spin of it, so comparing clock for clock against current Nvidia offerings we only see 5-7% increase in performance in performance, but the fact that they clock so high and so easily adds to the my second comment/reply about efficient use of caches and power consumption.

Given the current thermal output and cooler ability of 7970's 1.1Ghz stock clock seems entirely reasonable. But that probably would have created an artificial deficiency of cores that met their specs. Knowing them they are probably binning so that when the 7950 comes out they will have enough stock for it and for the next large batches of 7970, giving them more than enough time to get stock of other slightly defective cores for midrange and some low end cards, and also some pristine cores for use in the dual GPU variants to combat whatever Kepler is bringing.

 

[Crap Daddy]

If we are in the rumor zone I would suggest another read:

http://semiaccurate.com/2012/01/19/nvidia-kepler-vs-amd-gcn-has-a-clear-winner/

The notorious NVidia hater Charlie Demerjian either has gone mad or was hacked. Or he is right.

 

[gorg_graggel]

i think the current situation is quite amusing...

they keep telling people that the gk104 is their performance class part.
though it's being presumed to be within the same performance/power envelope as amd's current offering...which is considered high-end.
rumors say, that by q2 the high-end gk110 part will be out, which is gonna be even faster, because that's supposed to be the actual maxed out kepler with a higher tdp...

it's seems to me to be basically the same strategy they used with gf100 and it's refresh gf110, just a bit optimized in terms of naming schemes to make people believe their new gen is so good, that it's cut down part even beats the competitions high end part.

what i think is, that they knew they're gonna be late to the party and had to come up with that "story". if they'd not been late it would not be gk104 and gk110, but gk100 and gk110...or even just gk110 named gk100 and no refresh at all...
not that they necessarily changed their plans along the road, but had learned their lesson from the little disaster the first fermi chip has been and how much of a redemption the fermi refresh was afterwards. so they adapted and optimized that process into a more controlled strategy, because it turned out to be very successful...

well anyways, considering tahiti seems to have quite a bit of reserves in it's design, amd could also release something along the lines of gk110's tdp and, guess what, close the gap again (or even surpass it). actually there have been rumors that this might just happen, so...
i think they might not have stopped optimizing the design further after they released tahiti...

amd shouldn't have booted most of the bulldozer marketing team. they should have switched them with tahiti's.
in my perception tahiti's pre-release info seemed kinda modest compared to what nvidia is doing or what the bulldozer marketing tried to pull off with the "FX" brand (and failed).
being modest helps not being "bulldozed" (pun intended) by disappointed customers because of high expectations, but makes more people not buy your product initially (although it's very good), because the competition claims theirs will be better, so more people wait for it... high risk, high reward...kinda...

nvidia's marketing seems to know better how enthusiast's brains work, as lots of people seem to fall for it...
i know what you did there you dirty, little rascals!


of course, i have pulled all this out of my arse and don't claim any of it to be true...but i think i might be somewhat thinking in the right direction here...
it's all just a few "elaborated" (actually not) guesses (yeah, more of that!)

myself being a cheapskate/value type buyer and underdog lover, amd's strategy caters more to me. so, while i think the 7970 deserves to be priced higher than the former performance leader, i won't ever buy a card at its current price point again (like i did, back in the days...sweet gf2ultra ...set me back about 600€/1100 deutsche mark)...
so i kinda hope the new nvidia cards to be a bit faster than the amds, so they can drop to a price i'm willing to pay (about 350€)...

well, that post branched out more than intended, but well...there you go...

 

[AsRock]

Ah, the predicted "but wait, we have something too" response from Nvidia. I wonder how good Kepler is going to be, given the massive power consumption.

Yes gotta love that, we have some thing better but wait it's not ready yet.. And if it not done by nvidia it's done by a fan of nvidia and the shit happens with both sides.

Just to damage AMD sales by getting people hold on their money for a while longer which really how i see it don't really matter as AMD don't have enough 7900 cards to supply yet.

 

[Steevo]

This round I am willing to pay for performance, but not until I know for sure its what I want. I might get a green card in my rig. But I love to overclock red stuff.

 

[Benetanegia]

If we are in the rumor zone I would suggest another read:

http://semiaccurate.com/2012/01/19/nvidia-kepler-vs-amd-gcn-has-a-clear-winner/

The notorious NVidia hater Charlie Demerjian either has gone mad or was hacked. Or he is right.

2012... world... boom

 

[OneCool]

Smaller is faster

 

[overclocking101]

it would not surprise me at all if this is true. it would make sense that the first card would be a 560 replacement so that nvidia has time to make the higher end cards better. It's very common. and if the performance is what they say it is going to be ill be on the green team once again.

 

[EarthDog]

2012... world... boom

hahahahaha, I take back everything I said about that lieing ass tool.

 

[blibba]

256 bit bus means cheap to make.

This card, like GF104, is designed for price wars. They'll make a high-end behemoth too, don't worry

 

[MxPhenom 216]

Can't wait to fry my bacon and eggs for breakfast on this card. Sorry Nvidia, but I honestly don't see this card beating the HD7970

 

[Kaynar]

C'mon man! ATIs 7970 has twice the transistors count (2640M vs 4310M transistors - 6970 vs 7970) but has only 20% or little more perf vs last gen of cards so think twice! ATI failed to sqeeze the perf out of those transistors. So lets hope Nvida will not fail with his architecture

Remember they are using a new architecture on this card, so u cant compare the transistor count as a simple % increase... i'm sure their next cards (i.e. HD8000 series) will be a small revision of the HD7000 and will have another performance bump.

I have an HD7970, it scores P8380 in 3dmark11 standard benchmark, which I believe is 15% better than GTX580 and the GTX580 is another 10% or so better than HD6970. I won't be surprised is nVidia new cards are 5-10% faster than HD7970 and cost $100 more.

 

[blibba]

Can't wait to fry my bacon and eggs for breakfast on this card. Sorry Nvidia, but I honestly don't see this card beating the HD7970

I don't think it'll match it for multi-monitor support or idle power. But out and out performance, I don't think Nvidia have the balls to release anything that doesn't right now.

 

[Casecutter]

Read it again tt just says, "NVIDIA is gunning for the performance crown from AMD Radeon HD 7900 series with this chip"
If it sounds too good to be true... I see this as smoke.

The GK104 will provide a few "cherry picked" chips that AIB's will again turn into Uber OC "For the Win" units that might encroach into 7970 territory, but hardly going to have the loins share of offerings holding to that.

The fight will be against the 7950 at $400 MSRP; Nvidia will have more of O.C's that will best a 7950, but probably be more than $400, while minimal equipped reference clocks will be made to spare with a 7950. But as normal it will be price/equipment and this time almost every AMD AIB’s will come straight out of the gate, with nice coolers and OC's, while maintain within an MSRP of <$400.