Big Kepler news

[blckgrffn]

Yea, I wouldn't be surprised to see a 300 TDP.

Yep, I agree. Why not? It's to their advantage to go all in - and then charge for it.

A dual GTX 680 card is going to be their halo gaming card, not Big-K.

That's my prediction

Things would be a lot different today if that were the case. GK104's for $250 would be real nice. I would have 2 already.

Like the 4870?

Yeah, that works great for a bit but if you are in a position of strength all it does is erode the value of your brand.

See: All the whining about AMD 7xxx series parts (Guilty here, too...)

 

[Edrick]

Would you?

They can't keep them in stock at current ~$500. What do you think demand would be like at ~$250. Buying one would require an army of browser tabs at different e-tailers with a macro that cycled through the tabs hitting F5 and looking for stock.

Fair point, I would try real hard though

 

[Vesku]

7770 is closer to the 6850 than the 5770 in terms of DX 11 gaming. What we are seeing, I think, is that AMD's goal is to bring GPGPU computing to their whole product line (including integrated with CPU) while Nvidia is keeping their solution enterprise priced with a high margin.

I am curious to see what tweaks they make to larger die Kepler in order to make it a GPGPU successor to 580.

Edit: I wonder if the GPGPU unit will be much closer to a beefed up 580 than the 680.

Nvidia basically crippled 680s compute because gamers don't need it and they want to sell Teslas for $1000s. AMD on the other hand used the extra transistors from 28nm to beef up compute; for instance 7770 has 1.5 billion transistors vs 1.04 billion for 5770 yet performs about the same. That's a 50% 'compute tax'. I wonder if that's viable since it means Nvidia can cut compute and still perform the same on maybe 70% of the die. Of course, Fusion needs compute to take off....

 

[digitaldurandal]

7770 is closer to the 6850 than the 5770 in terms of DX 11 gaming.

Not sure if serious... or trolling.

 

[KompuKare]

7770 is closer to the 6850 than the 5770 in terms of DX 11 gaming.

I was going more by the average scores which various sites do. However 7750 is the same die and it performs below the 5770. Anyway, my point was more along the lines that a 1.5 billion transistor budget is nearly 50% on top of 5770 or looked an other way the same as Barts (6850/70). So if they had shrunk Barts to 28nm and pushed it's clock even if with a 128-bit bus it would probably have been far faster.

What we are seeing, I think, is that AMD's goal is to bring GPGPU computing to their whole product line (including integrated with CPU) while Nvidia is keeping their solution enterprise priced with a high margin.

Indeed, that's probably the same reason they didn't shrink Thuban / Stars since the BD modules are designed to be easily dropped into the same die as their GPUs and are meant to use the GPU rather x87 FPU.

 

[bunnyfubbles]

these numbers aren't adding up

if true GK110 will be a much larger departure from GF110 than GK104 was from GF114

 

[SickBeast]

these numbers aren't adding up

if true GK110 will be a much larger departure from GF110 than GK104 was from GF114

I'm thinking they dedicated more of the GPU to GPGPU tasks and lowered the clockspeed, which is why they're saying 25% faster as opposed to 50% based on the specs.

 

[boxleitnerb]

I expect significantly more performance than just +20-25%. Rumors put the GK110 at 3072, not 2304 shaders, albeit lower clocks, maybe 800-850.

 

[SickBeast]

There are just too many unknown factors that make Kepler an unlikely release in September IMO. 22nm yields have been bad, and this chip is huge. TSMC may not be able to manufacture it yet.

 

[Rvenger]

I bet there is not much more performance to squeeze out of GK110, hence why they used GK104. GK110 was designed for compute power not just raw gaming performance. If Intel left AMD in the dust then Nvidia will leave AMD in the dust. Plain and simple.

 

[BallaTheFeared]

I fully expect to lose some performance due to compute eating up perf/watt, how much is unknown because unlike GF104/114 vs GF100/110 GK104 has even less compute than GF104/114.

The real issue for me is the overclocking, the 680 isn't anything I'd care to own for the simple fact that it is so limited in what you can do with it. Hopefully BigK addressed what I feel is a major blunder on an enthusiast priced card.

 

[KompuKare]

The real issue for me is the overclocking, the 680 isn't anything I'd care to own for the simple fact that it is so limited in what you can do with it.

I have a theory about that but around here the words GK104, mid-range, stock overclocked and 'we saw where 7970 was first' don't go down too well...

 

[bunnyfubbles]

I'm thinking they dedicated more of the GPU to GPGPU tasks and lowered the clockspeed, which is why they're saying 25% faster as opposed to 50% based on the specs.

still not adding up, GF110 was much faster than 20-25% over GF114 despite specs that were only ~25% better, even if GK104 is more gaming efficient at the cost of compute these numbers suggest GK110 isnt improving at all and thus would have to be an absolute monster compute chip by more than just average generational improvement to make up for it.

I fully expect to lose some performance due to compute eating up perf/watt, how much is unknown because unlike GF104/114 vs GF100/110 GK104 has even less compute than GF104/114.

The real issue for me is the overclocking, the 680 isn't anything I'd care to own for the simple fact that it is so limited in what you can do with it. Hopefully BigK addressed what I feel is a major blunder on an enthusiast priced card.

25% OC on my 680 is just as good if not better than anything I ever got out of my 4 Fermi cards, including two 580s, granted that's just air cooling

 

[Lepton87]

I don't think it will cost more than 600$ because then it would have to compete with dual tahiti and maybe dual GK114 and it will be much slower than either of those cards.

 

[aaksheytalwar]

Not really, it can cost $650 as well though it may not, dual gpus will be upwards of. $850

 

[bryanW1995]

not too surprising really. I see BK as having to cater for their Tesla range so it will concentrate on compute. While 680 has 1/24 DP performance vs SP, this will have to do a lot better because ATM for some compute stuff, 7750/7770 beats the 680. Wiki says the 680 has 3090 GFLOPS while 580 had 1580. I assume that's SP but for Tesla DP is a must.

So far this generation seems to have gone thus: Nvidia went for SP and gaming performance while AMD went for DP and compute.

Nvidia basically crippled 680s compute because gamers don't need it and they want to sell Teslas for $1000s. AMD on the other hand used the extra transistors from 28nm to beef up compute; for instance 7770 has 1.5 billion transistors vs 1.04 billion for 5770 yet performs about the same. That's a 50% 'compute tax'. I wonder if that's viable since it means Nvidia can cut compute and still perform the same on maybe 70% of the die. Of course, Fusion needs compute to take off....

That's true as far as it goes, but it ignores the main point: NV used what was supposed to be their gtx 560ti replacement and clocked the crap out of it b/c they didn't want another iteration of the fermi I launch. Now, instead of huge criticism for forcing bigK out before it was ready, they instead get huge kudos for bringing out the best/fastest gpu for consumer gaming on a smaller die than AMD is using. Of course, AMD helped by offering up such weak performance.

 

[Smoblikat]

Everyone... I mean rational people already know the 680 is a mid range card. Dropping it to 399.99 would make it an even better steal and placing the 685 at 499.99 would be incredible. Ofcourse this is all speculation.

I started a thread about this very fact, i couldnt have put my flame suit on fast enough. People berated me for even CONSIDERING the possibility that the 680 was intended for the mid range market.

 

[Vesku]

Can't see why people would be so angry over the discussion. GTX 680 is lean and tuned for pure video game performance, to think that it was meant to be the 28nm GTX 580 requires just as much interpretation as thinking it's the 28nm 660/670 repurposed.

To say "but it's named the 680 so it is the 680" is missing the point entirely. Compare this to 7970 default clocks discussion about whether AMD was milking initial 28nm dies, was there a ground swell of "but those are the clocks so that's what the clocks are" not that I saw. Because it would be silly and missing the point.

Might as well just have a sticky saying: "The big die GTX 600 series will be what it will be, if it shows up at all." I guess it is the best kind of correct...

I started a thread about this very fact, i couldnt have put my flame suit on fast enough. People berated me for even CONSIDERING the possibility that the 680 was intended for the mid range market.

 

[KompuKare]

That's true as far as it goes, but it ignores the main point: NV used what was supposed to be their gtx 560ti replacement and clocked the crap out of it b/c they didn't want another iteration of the fermi I launch. Now, instead of huge criticism for forcing bigK out before it was ready, they instead get huge kudos for bringing out the best/fastest gpu for consumer gaming on a smaller die than AMD is using. Of course, AMD helped by offering up such weak performance.

I do wonder if that was their intention all along or (and this my theory) they only did that once they either knew the concrete or leaked 7970 performance.

Well yes, Tahiti 365mm² vs 294mm² is quite a bit. Can't help but think though that the Pitcairn at 212mm² offers pretty good per/watt/die-size the problem is AMD were pretty conservative with that design (although they did clock it 1GHz). If they had released three months later it's entirely possible they might have tweaked both Tahiti and Pitcairn a bit more. Maybe in the 28nm refresh, AMD will run with the Pitcairn design, add a few more unit to get a ~300mm² die, use faster RAM and have more of a difference between it and the bigger Tahiti design. At this stage I think it's a bit late for AMD to remove Compute features for their next 28nm design (unless their design is very very modular) and they may not want to because of Fusion.

 

[epidemis]

http://www.legitreviews.com/news/12803/

TBH I'm disappointed if it's true. I was expecting more than 20-25% faster than the GTX 680.

Incidentally the power consumption is up 25%..

 

[Keysplayr]

Bigger leap of cores?
384 -> 512
1536 -> 2048
Sure, the number is bigger, but it's still the exact same percentage increase, 33%.

Where did you get 2048 from?
I'm seeing 2304 which would put the increase in cores from GK104 to Big Kepler around 50% increase.
768 + 768 = 1536
768 + 768 + 768 = 2304

 

[toyota]

Uh.

Is it just me or 50% cores and double the memory bandwidth should bring more than 20-25% increase?

yeah that makes no sense at all. that would have to be the poorest scaling architecture in history if true.

really I do not see the point in going to 512 bit bus with the 384 bit bus would probably get the job done.

 

[Akantus]

With 50% more cores, 512bit bus + GPGPU stuff and only 250W TDP, it's gonna be clocked quite a lot lower than 680. That's probably why only 20-25% improvement...

But I think with some nice aftermarket designs and OC, it has potencial to go to 50% improvement. Of course at 300+ W TDP :twisted:

But that's just my guess

 

[Akantus]

really I do not see the point in going to 512 bit bus with the 384 bit bus would probably get the job done.

384 bit would probably be fine for games, but I think GPGPU could take advantage of 512 bit bus. And this card is GPGPU first gaming second...

 

[iCyborg]

yeah that makes no sense at all. that would have to be the poorest scaling architecture in history if true.

really I do not see the point in going to 512 bit bus with the 384 bit bus would probably get the job done.

7970 has 60% more cores and ~70% more mem bandwidth than 7870, but is only ~30-35% faster on average.
nV cranked the clocks on 680 high to beat 7970...