Nvidia BigK GK110 Kepler Speculation Thread

[jimbo75]

But like silent_guy points out, things would have been completely different if AMD decided to release say the GHz version as the original 7970. 30~40% would have given it a more receptive welcome and sort of justify their abandonment of their small die/pricing strategy infavor of the old model.

I agree - the graphics division's worst mistake in the past 5 years was releasing the 7970 and 7950 with clocks that were too low. While I'm sure it made sense then, history shows it was the wrong decision.

Maybe the real fault was releasing so early while the process was still immature. Maybe that's why they are holding off now.

 

[Psycho]

Wait a minute, this card has ~15% higher base core clocks than a K20X and comes with faster RAM/more bandwidth while only bumping TDP by 6% and costing significantly less.

TDP and geforce-TDP are 2 different things..
Expect it to consume more than 15% more, or throttle, when running the same workload..

Even if it got a baseclock, which it will mostly be above in games, it will still throttle at certain workloads, while the tesla tdp is probably less capped.

 

[A1xLLcqAgt0qc2RyMz0y]

Yes, at least for people who don't care too much about ECC.

But it's strange, because GK110 is supposed to have 1/3 DP rate, which is not what these numbers show.

Maybe DP can only run at the base frequency whereas SP can run at the boost frequency

 

[ninelven]

That's irrelevant...

No, it is not. But I digress...

 

[Alexko]

Maybe DP can only run at the base frequency whereas SP can run at the boost frequency

The difference wouldn't be that large, Boost is less than 5%.

 

[trinibwoy]

Not that I disagree with your analysis but I recall a large number of people complaining bitterly about the 7970 which had a ~30% increase for around the same price as the 580.

It's not quite the same scenario. The 7970 was launched to much fanfare on a new process and on a new architecture. Neither applies to Titan so obviously it won't be held to the same standard or be subjected to the same level of scrutiny.

 

[jimbo75]

It's not quite the same scenario. The 7970 was launched to much fanfare on a new process and on a new architecture. Neither applies to Titan so obviously it won't be held to the same standard or be subjected to the same level of scrutiny.

To be honest a new GPU architecture generally means that it'll be a while before drivers are good enough (at least in AMD's case). The 7-series is the perfect example of it.

If anything this points to Titan being more disappointing as previous drivers should be at least up to speed?

If you think about it in other ways, say compare the 7970 GHz to the 6970 - the 7970 GHz is some 60-80% faster and was released only 18 months later. Titan is going to be some 95%-100% faster than the 580 some 2+ years later. That's pretty similar stats - die sizes remain fairly constant, change in process etc all factored in (that's the point I was trying to make to Ninelven).

There's really not a lot of difference between both companies improvement, but the price doesn't appear to fit the cards performance in Nvidia's case.

 

[silent_guy]

If anything this points to Titan being more disappointing as previous drivers should be at least up to speed?

Yes, jimbo. The Titan is utterly disappointing. The 7970 treatment was a terrible injustice. And you and I are the only ones to realize this in an unfair and biased world.

Good enough? Will it make you stop this line of 'argumentation'? What do I need to add?

 

[jimbo75]

Yes, jimbo. The Titan is utterly disappointing. The 7970 treatment was a terrible injustice. And you and I are the only ones to realize this in an unfair and biased world.

Good enough? Will it make you stop this line of 'argumentation'? What do I need to add?

Nothing, I'm glad you agree. On the other hand it's not about you (this time) sorry to say!

The inconsistencies on each side need to be brought to the forefront. I was a pretty damning critic of the 7970 and if Titan's performance and price is consistent with what we're being led to believe, then I sure hope to see some damning criticism from the green teams fans as well - not some apologetic pish that they can charge what they want for having the fastest card.

 

[silent_guy]

Like I said: based on area, the Titan performs pretty much as expected.

 

[RedVi]

Yes, GTX Titan should have very good performance per watt compared to all 7970 GHz Edition cards.

Note that in some instances the performance per watt of the 7970 GHz Edition appears to actually decrease when overclocked. For instance, when Anandtech tested the 7970 GHz Edition at stock vs. overclocked operating frequencies using Metro 2033, overclocking resulted in a 9.6% increase in performance (http://images.anandtech.com/graphs/graph6025/47611.png) but a 13.1% increase in load power consumption (http://images.anandtech.com/graphs/graph6025/47618.png). On the other hand, when Anandtech tested the GTX 680 Classified at stock vs. overclocked operating frequencies using Metro 2033, overclocking resulted in a 8.9% increase in performance but only a 4.8% increase in load power consumption. So it will be interesting to see how GTX Titan fares in this area too. Also note that overvolting (as opposed to overclocking) was not very useful on the GTX 680 Classified, since overvolting + overclocking resulted in a very small performance increase vs. overclocking alone. Since the GTX Titan clearly supports overvolting, the assumption is that overvolting will achieve much better results on the GTX Titan than on the GTX 680.

Tahiti is a bulky compute orientated GPU though, GK104 is more like a bigger Pitcairn. That's like saying GF114 scaled in clock/watt better than Cayman so GF110 must also. Clearly Titan has lower clocks for a reason, if they could commonly scale to higher clocks with a better than 1:1 performance/watt increase then surely they would have clocked it higher. Expecting it to scale like one particular GK104 does is very presumptuous.

 

[ninelven]

The inconsistencies on each side need...

Windmills are tough opponents..

 

[mczak]

Also note that overvolting (as opposed to overclocking) was not very useful on the GTX 680 Classified, since overvolting + overclocking resulted in a very small performance increase vs. overclocking alone. Since the GTX Titan clearly supports overvolting, the assumption is that overvolting will achieve much better results on the GTX Titan than on the GTX 680.

I wouldn't bet on it. I think to get the seemingly quite good power consumption numbers voltage is probably quite low. So yes overvolting could get you into 1Ghz range but I don't think the card would allow the huge power draw increase. Though since it's a very low-volume part I guess there's some hope.
The numbers otherwise look ok to me, the thing does what it's supposed to do (be the fastest single chip card), but obviously if you want perf/$ look elsewhere. If it has 1/3 DP at least a few people will really like that (which might otherwise have got 7970...).

 

[gamervivek]

Windmills are tough opponents..

where art thou dammitcenia?

 

[trinibwoy]

To be honest a new GPU architecture generally means that it'll be a while before drivers are good enough (at least in AMD's case). The 7-series is the perfect example of it.

That may be true but in reality does very little to temper expectations.

There's really not a lot of difference between both companies improvement, but the price doesn't appear to fit the cards performance in Nvidia's case.

The competitive landscape is very different so you're trying to draw a parallel that really doesn't exist. There are no grand expectations for Titan since the underlying architecture is old news. Great disappointment only follows grand expectations.

Also, there is no new manufacturing process around the corner to render it obsolete soon after launch. Are you expecting AMD to release a 28nm part with sufficient performance to force a price drop on Titan?

 

[Ailuros]

Good thing that people who will buy it have no need for such justification. 80% more expensive for 40% more performance is not as ludicrous as some try to make it seem. It's a relative price/perf bargain compared to high end CPUs for example.

A minority report of a minority report then.

Also, if Charlie is right and "GK114" is just slightly bigger/faster than GK104 then "Titan" will be king for a long time.

It would be highly "convenient" either way to not release a GK104 successor all that soon. Also considering that it doesn't seem that Curacao and co. won't appear before Q3 either (for which Charlie projected an analogue performance increase), there's a fine chance to go for limited wafer runs keep the costs at a minimum and pump up the price as much as possible for an uber-expensive halo product to win benchmarks in reviews in the meantime. That way they don't need to worry about any serious price reductions on GK104 and co either and continue the high margin/low volume trend we're seeing for roughly a year now.

I'm all eyes to read some educated speculative math that really justifies an 80% MSRP increase over the GTX580. There's a reason why I asked pages ago instead of financial results for volume persentages on an annual basis; if we would have the chance to have volumes per die area statistics it would be even more interesting. I've no doubt that 28nm manufacturing might be up to a reasonable persentage more expensive than 40nm manufacturing used to be. Given that we already have a year of 28nm GPUs on our back, differences should be smaller.

Don't think that I'll care if they'll continue to alienate an increasing amount of gamers with those kind of strategies.

Also, there is no new manufacturing process around the corner to render it obsolete soon after launch. Are you expecting AMD to release a 28nm part with sufficient performance to force a price drop on Titan?

Not before Q3 I guess. Things are inevitably slowing down due to processes getting increasingly problematic, but it's not like IHVs are actually doing anything against it; au contraire it look rather they're assisting the whole process to slow down even further. Granted of course they're utmost priority is to protect their interests first above all, but if you think that users are in their majority just blind idiots in the end you might want to check again.

 

[jimbo75]

The competitive landscape is very different so you're trying to draw a parallel that really doesn't exist. There are no grand expectations for Titan since the underlying architecture is old news. Great disappointment only follows grand expectations.

I dunno but it seems to me like Titan has been played up more than Tahiti ever was, maybe because of the success of the 680.

Also, there is no new manufacturing process around the corner to render it obsolete soon after launch. Are you expecting AMD to release a 28nm part with sufficient performance to force a price drop on Titan?

By the end of this year or the early part of next year, yes. To be frank I think the GHz edition will be close enough that Titan will look terrible value anyway.

 

[UniversalTruth]

Sure, it is poor value. I would prefer to buy 3x 350€ Radeon 7970s instead of one 1 k€ Titan.

Not to mention that even with a single 7970 you would do just fine.

 

[lanek]

That may be true but in reality does very little to temper expectations.



The competitive landscape is very different so you're trying to draw a parallel that really doesn't exist. There are no grand expectations for Titan since the underlying architecture is old news. Great disappointment only follows grand expectations.

Also, there is no new manufacturing process around the corner to render it obsolete soon after launch. Are you expecting AMD to release a 28nm part with sufficient performance to force a price drop on Titan?

Basically we was allready know +- where was the performance, the only question was the clock speed, for get the details

whatever is / was the expectations about Titan, whatever is the performance on final, this will not disappoint anyone. Its not really a factor when it come from Nvidia. They take back the crown, and thats only this who will count. The same who was saying 2GB is enough will now pray the 6GB and whatever is the performance in addition, will just take it as the Faster card. Most who will pray it will never buy the card anyway..

I believe Nvidia count to drain the sell of actual 600 series even more with Titan.. if they not release a 700 series soon based on GK114, this will stay the affordable cards for a while on their brands.
High price of it, ensure the 680-670 price dont have to drop much.

I and many other dont have problem for buy one GPU at 650-700$ if it is the top high end available. But most gamers dont even look at cards over the 350$ ( 7870, GTX 660 (TI )), GTX560 etc ).


About AMD, well in reality im not so sure they will need so much for close the gap ( on paper, 20% shader more, same clock speed, they break the 5Tflops ), but they should better wait to see if Nvidia release a GTX 700 series first.
If they move before Nvidia, the risk is to see Nvidia release aditional cards based on GK110 between GK114 and the Titan. And Im sure Nvidia is allready looking at a second version: 3Gb, one SMX less, 80-100$ less.

With a so high priced big GK110 Titan cards and the next GK114, Nvidia have all the latitude for fill any performances or prices holes against their competitor.

 

[trinibwoy]

I dunno but it seems to me like Titan has been played up more than Tahiti ever was, maybe because of the success of the 680.

Might just be a matter of perspective then. From where I stand Titan is a known quantity.

By the end of this year or the early part of next year, yes. To be frank I think the GHz edition will be close enough that Titan will look terrible value anyway.

A year is a long time away. Certainly too far away to have any bearing on a card launching this week....

If nVidia turns around and releases an even faster GK110 based part in a few months I'll admit it's shady but that's unlikely.