Nvidia still releasing the 7 series this year?

[spoffle]

My thoughts are on the next release from AMD being in Q4 of 2013 or even Q1 of 2014 and then not releasing another top card for at least 18 months. I do expect the 8xxx series to nip at Titans heels but not quite make it.

I am purely guessing though and could be miles wrong.

The AMD and nVidia release schedule is tied to the availability of smaller fab processes ultimately.

I don't think availability of high end 20nm GPUs is going to take 18 months, and the drop to 20nm from 28nm is about the same as it was going from 40nm to 28nm.

What's interesting is how much performance AMD managed to get from going to 28nm from 40nm. The 7970GE is about 2x the performance of the 6970, using 33% more stream shaders, on a GPU that's about 12% smaller. Going from 40nm to 28nm means about 30% more space.

The reason I mention that is because if AMD were to straight move a Tahiti GPU to 20nm, you'd get a 7970 performance GPU that's only slightly bigger than a 7870 GPU, with a bit more power consumption.

Basically, the move from 28nm to 20nm means there's going to be a load of room for a load more performance, and whilst it's guess work, there's really no reason to expect it'll take so long for the GTX Titan to be overtaken by a faster single GPU card.

You are entitled to your opinion, the same as I am mine. You are more than welcome to tell me I am wrong if Nvidia or AMD surpass the Titan by March of 2015.

You seem to think that the 8 series will beat the Titan by your way of thinking?

Whilst you are indeed entitled to your opinion, I think it's fair to say that your opinion on it isn't really an informed one.

As looking at the technical aspects of moving from 28nm to 20nm, there's absolutely no reason to think AMD's top 20nm GPU is not going to be faster than the GTX Titan.

If you look over the history of AMD's GPUs, more often than not, they double performance using 30% more GPU space.

3870 > 4870 more than doubled performance on the same process, using 30% more die space.

4870 > 5870, doubled performance using 30% more space on the same process.

6900s don't really count as they are like the GTX5 series, an efficiency refresh on the same process.

5870/6970 > 7970, dropped to 28nm giving roughly 30% more space to play with, doubled performance using about 12% less space than the 6970.

So as above, going from 28nm to 20, they could increase performance quite a lot, and even more if they went for a slightly larger GPU as they did with the 6970s at about 390mm² compared to about 350mm² for the 7970.

 

[andybird123]

if 20nm parts were less than 18 months away, there would be little to no point in releasing more 28nm parts later in the year like both NV and AMD are planning to do

I can see 28nm refresh parts coming close to or matching Titan, I can't see them clearly surpassing it

I hope Titan's do drop in price, so that I can get a 4th

 

[Rroff]

The reason I mention that is because if AMD were to straight move a Tahiti GPU to 20nm, you'd get a 7970 performance GPU that's only slightly bigger than a 7870 GPU, with a bit more power consumption.

Basically, the move from 28nm to 20nm means there's going to be a load of room for a load more performance, and whilst it's guess work, there's really no reason to expect it'll take so long for the GTX Titan to be overtaken by a faster single GPU card.

Unfortunatly for the move from 28nm its not that simple, you can't do a direct shrink from 28nm to 20nm theres several fundamental differences in the lithography which is likely to negate a lot of the room for more performance atleast in the first generation of products using it.

EDIT: I wouldn't be at all suprised if AMD hold off the 8000 series until 20nm and do another 28nm refresh under the 7000 series in the interim, not sure if people are just getting confused with the 7790 and thinking they are reading 7990 but I've seen a few rumours of an "ultra" type single core card to replace the 7970.

 

[spoffle]

Unfortunatly for the move from 28nm its not that simple, you can't do a direct shrink from 28nm to 20nm theres several fundamental differences in the lithography which is likely to negate a lot of the room for more performance atleast in the first generation of products using it.

It was an example of what going from 28nm to 20nm will mean in terms of die size, performance and power consumption.

 

[Gregster]

You have nothing to back your claims, the same as I don't. The only rumours floating about ar that the 8 series will still be on 28nm and Volcanic islands will be on 20nm.

Whilst you act out your superiority complex and know it all fashion, you have absolutley no evidence of the claims you have just made. You have not provided any informed evidence of you being remotely correct.

http://www.kitguru.net/components/g...adeon-hd-8970-and-8950-launch-plans-revealed/

Going from that slide, you can see that the 8970 is on 28nm (not sure why you keep on about 20nm?).

Whilst I thankyou for allowing me to have an opinion, bring some evidence next time you wish to call me delusional or uninformed

Edit:

What's interesting is how much performance AMD managed to get from going to 28nm from 40nm. The 7970GE is about 2x the performance of the 6970, using 33% more stream shaders, on a GPU that's about 12% smaller. Going from 40nm to 28nm means about 30% more space.

Not sure where you are getting your figures but you are massively wrong stating that the 7970GE is ~ 2X the performance of the 6970

When we look at Anandtech, you can see that your guesstimate of 2X the performance is massively delusional.

http://www.anandtech.com/bench/Product/509?vs=618

Just using the 1920x1200 benchs:

7970GE is faster than the 6970 by 50% in Crysis Warhead
7970GE is faster than the 6970 by 49% in Metro2033

Just 2 examples of how wrong that is and the increases are far less with higher resolutions.

 

[spoffle]

How do you know I have no evidence? Because you have no idea about die sizes and the performance gains from going between them?

I'm talking numbers, they're all there for you to check if you're so bothered about insisting that I'm not providing proof, it's like you're asking me to provide evidence that 2x2 = 4, whilst ignoring the fact that I'm talking about the calculation.

They're facts, and you can't argue with that. You also need to get over the "superiority complex" thing as well. You readily state that you don't understand the technical aspects, so why get bent out of shape when I point out that your guess work isn't based on the technical aspects, but more on how you want your Titan cards to be at the top for longer?

Additionally, you do know that the chart you're talking about is just a reference from Wikipedia based on rumours right from 6 months ago, right? The Wikipedia page that has since been changed. I'm not saying it's wrong, but it seems like you think that chart is some secret leak that people weren't supposed to see. The article also states that these cards are to be showed off during CeBIT 2013, which ended nearly 3 weeks ago...

*** Then ignore them spoffle, don't make it personal by making personal remarks **

 

[Gregster]

I checked your numbers and they don't add up. No point debating if you make figures up.

 

[Rroff]

3870->4870 is pretty much twice the performance, 4870->5870 is around +70%, 6970 to 7970GHz on latest drivers is about +60% on average, around +40% on release. Most of those big generation jumps on AMD and nVidia have been around +70% average performance increase high end->high end GPU typically.

 

[spoffle]

You cherry picked 2 examples from the Anandtech GPU benches that are using old out of date drivers for one.

You also haven't checked my numbers and they don't add up, because they do add up. Maybe you're just adding them up wrong?

 

[Diagro]

I also think the next gen will be 30%+ faster than the 7970, Which in turn will match or be faster than the Titan, Simple.

 

[spoffle]

3870->4870 is pretty much twice the performance, 4870->5870 is around +70%, 6970 to 7970GHz on latest drivers is about +60% on average, around +40% on release. Most of those big generation jumps on AMD and nVidia have been around +70% average performance increase high end->high end GPU typically.

It depends on what games you look at, some games like clock speed more than shader counts, and some prefer shader counts. It's not as simple as Gregster's intentionally flawed example.

We all know that on release, the 7000 series wasn't a massive jump on the 6, but they have made quite a lot in the way of gains.

 

[spoffle]

I also think the next gen will be 30%+ faster than the 7970, Which in turn will match or be faster than the Titan, Simple.

What I was getting at earlier was more that if the next gen is 28-20nm, there's absolutely no reason that the jump will be that small that we see a top end card that's still slower than Titan.

 

[Diagro]

What I was getting at earlier was more that if the next gen is 28-20nm, there's absolutely no reason that the jump will be that small that we see a top end card that's still slower than Titan.

I agree, I would also expect the Next Gen from nvidia to be faster than the titan.

Anyone who seems to think otherwise is ignoring all previous generation performance jumps or is an owner of a Titan is is trying to justify the outlay of money

 

[spoffle]

I agree, I would also expect the Next Gen from nvidia to be faster than the titan.

Anyone who seems to think otherwise is ignoring all previous generation performance jumps or is an owner of a Titan is is trying to justify the outlay of money

Well exactly, that was my point really, but yet again some one who doesn't understand the technical aspects, tries to argue them.

I also think that nVidia's next gen will outperform the GTX Titan.

 

[andybird123]

How do you know I have no evidence? Because you have no idea about die sizes and the performance gains from going between them?

I'm talking numbers, they're all there for you to check if you're so bothered about insisting that I'm not providing proof, it's like you're asking me to provide evidence that 2x2 = 4, whilst ignoring the fact that I'm talking about the calculation.

They're facts, and you can't argue with that. You also need to get over the "superiority complex" thing as well. You readily state that you don't understand the technical aspects, so why get bent out of shape when I point out that your guess work isn't based on the technical aspects, but more on how you want your Titan cards to be at the top for longer?

Additionally, you do know that the chart you're talking about is just a reference from Wikipedia based on rumours right from 6 months ago, right? The Wikipedia page that has since been changed. I'm not saying it's wrong, but it seems like you think that chart is some secret leak that people weren't supposed to see. The article also states that these cards are to be showed off during CeBIT 2013, which ended nearly 3 weeks ago...

Oh lawd, I don't know why I even bother with you, because you just go in to baby mode when you're disagreed with.

Looking at lots of 7970 reviews, even on the beloved 12.11 drivers, and 7970's are not getting 2x the fps of 6970... shouting at someone that they are wrong and can't do basic maths when you make glaring mistakes like that is in itself pretty babyish...

 

[Diagro]

* mod edit - nothing to do with the topic **

Gregster, You actually think the TITAN wont be beaten by 2015? Going by past trends of generation performance jumps it will be beaten or at the very least matched.

 

[Duff-Man]

I also think the notion that it's going to take 18 months for something to come out that's faster than a Titan is incredibly wishful thinking.

That's basically suggesting that it's going to take AMD nearly 3 years to come out with a chip that's only 20-30% faster 7970 since the 7900s have been out for about 14/15 months now.

That's almost to the level of delusional when it comes to wishful thinking if I'm being blunt about it.

GPUs are parallel computing devices (much unlike the CPU). Given this, performance is largely determined by the number of transistors you can cram into a die.

This leads to large performance jumps when the manufacturing process shrinks; going from 40nm to 28nm allows more than twice as many transistors to be crammed into the same area, and the drop to 20nm will be similar. TSMCs 20nm process is about 18 months away from mass production, so this is when we will see "titan beating" cards.

Inter-process refreshes offer much smaller gains, precisely because they are restricted to using the same size transistors. Yes, the engineers can usually eek out a few percent extra performance by improving the pipeline efficiency, and yes sometimes the clockspeed is bumped up a little. But the improvements are typically small - 15-25% is typically what we have seen from these inter-process refreshes over the past decade.

There is no reason to believe that this time around will be any different. Yes, AMD or Nvidia COULD make a faster GPU than the Titan, on 28nm, by using a massive die size. But the larger dies are much more complex, less power efficient, offer poorer yields, and take longer to "iron out the bugs". AMD have not gone for the "big die" approach since the 2900xt, and have shown no intention of doing so in the near future. Nvidia already has their "big chip" - the GK110 on which the Titan is based. The only reason that the Titan exists is due to the Tesla K20 (a card designed for scientific computing). Nvidia are not replacing the K20 until Maxwell, so there will be no "Titan buster" coming from them either.

So no, it isn't "delusional" or "wishful thinking" as you suggest, it's simply a case of cold hard logic. Historically, technically and realistically there will be no "Titan busters" on the market until 20nm cards arrive, and that isn't going to be for *at least* 12 months, realistically more like 18. The high-end 28nm refresh cards will get close to Titan performance, but it's highly unlikely to be "beaten" in terms of overall performance until 20nm rolls in.

 

[Duff-Man]

4870 > 5870, doubled performance using 30% more space on the same process.

No... the 4870 was built on a 55nm process, whereas the 5870 was on 40nm. 40nm allows 89% higher transistor density than 55nm, and the 5870 had 30% extra die size. So it was "quite a lot more GPU". Transistor count for the 4870 was 956 million, for the 5870 it was 2150 million. So it used more than double the transistors to gain roughly double the performance. A fairly typical die-shrink situation.

5870/6970 > 7970, dropped to 28nm giving roughly 30% more space to play with, doubled performance using about 12% less space than the 6970.

Remember you're dealing with AREA, not length. So, 40nm -> 28nm increases transistor density by (40/28)^2 = 2.04x (i.e. 104% increase).

Transistor count for the 7970 is 4.3Bn, vs 2.6Bn for 6970 and 2.1Bn for the 5870. Again, we see performance increasing roughly in line with transistor count.

A fair proportion of the Titan is devoted to compute-oriented features, so it would theoretically be possible to match its performance using a streamlined "gaming-centric" design by using fewer than 7Bn transistors. But you're still looking at a 30-40%% increase in transistor count, and therefore die size, to match the Titan's performance on 28nm. That really isn't likely to happen.

 

[Rroff]

snip

But but hard maths is no fun :| but your right, as an outlier and its not really been AMD's way of doing things of late they could put out a ~470mm2 core bringing them almost to Titan performance but I can't see them going larger than that and even thats a stretch. In the cold light of logic a Titan buster almost certainly isn't on the cards any time soon.

 

[Duff-Man]

But but hard maths is no fun :| but your right, as an outlier and its not really been AMD's way of doing things of late they could put out a ~470mm2 core bringing them almost to Titan performance but I can't see them going larger than that and even thats a stretch. In the cold light of logic a Titan buster almost certainly isn't on the cards any time soon.

Yes - a ~500mm^2 version of the 7970 design, on 28nm, should be able to match the Titan. But I don't see AMD producing a GPU of this size. Even ATIs largest GPU, the 2900xt, was 'only' 408mm^2, and ATI/AMD have been very much reluctant to produce large GPUs. I agree that it's unlikely that we will see such a chip. Realistically, we should expect the 8970 to be either 2304 or, at a push, 2560 stream processors. That is unlikely to be enough to beat the Titan overall, though it should be fairly close behind.

For those that are interested, here is a summary of the "vital stats" for modern high-end GPUs that I put together: