AMD Cayman info (or rumor)

[tajoh111]

1 Question.

1. Would AMD switch to vliw4 if there was only negative changes?
No.

So, at worst case, vliw4 is equivalent to Barts in performance.

Would 71% shaders extra to a barts architecture plus double the frame buffer and a new tess engine provide a big performance increase?

There's your cream.

There has to be some negatives, there are rarely changes that only result in positives. Increased power consumption is already one.

I think your assuming everything is shader limited and there will be a linear increase in performance with shader increase. This has been proven wrong almost entirely this generation.

The increase from a 5870 to a 4890(same clocks) was between 30-50%
Or even the gtx 280 to 480 is around 30-50%.

http://www.tweaktown.com/articles/29...on/index4.html

This was a best case scenario because AMD got to double everything, texture units, ROPS, Shaders.

http://www.anandtech.com/show/2841/17

Drivers might have brought performance up 5% but it also brought up 4890 and 4870 performance as well.

AMD this time is going to at best increase performance by 50% because there is only a 50% increase in real shaders. There is a big but this time around.

AMD is not going to double ROPS and texture units this time around; because of this, AMD will start seeing drops off from that 50%.

Even the wildest optimist on this board beside yourself, doesn't believe an 80% increase will happen because this is the same node.

Much of the purpose of changing to a new architecture again, is trying to make those spec gains linear again. Because if their was only a 1920 shaders difference(using the same technology), their might be a 5% performance increase between the 5870 and 6970, because the 5850 and 5870 perform the same at the same clocks and thus shows, shaders are encountering another bottleneck in the architecture. Changing to a new architecture is going to help the generation pick up some gains again, but not this perfectly linear 80 percent your thinking of. Barts XT performs as well as it is, because it has the ideal configuration(encounters the least amount of bottlenecks) to get maximized performance out of the r600 - r800 architecture. AMD is not going to get this ideal architecture off the bat, as it took AMD 3 and a half years to get there.

BTW, Barts xt is clocked at 900 and not 725mhz.

[Flying duck]

Snip...

Actually you have just highlighted that the 5870's are actually not very efficient and that there should be lot's of room for improvement with the new gen.

[tajoh111]

Actually you have just highlighted that the 5870's are actually not very efficient and that there should be lot's of room for improvement with the new gen.

I agree to some extent. However, although there is a lot of room for improvement, getting 50% extra performance is already showing that improvement. If AMD was going to bump the specs as is(1920 shaders), with the same architecture, we would be seeing a tiny performance boost(5-10%). Not this 30-45% I am expecting.

The only way you would see a 80 percent improvement jump is if you had perfect drivers and a game which entirely was built for the 6970 so no resources were wasted.

[mAJORD]

There has to be some negatives, there are rarely changes that only result in positives. Increased power consumption is already one.

I think your assuming everything is shader limited and there will be a linear increase in performance with shader increase. This has been proven wrong almost entirely this generation.

The increase from a 5870 to a 4890(same clocks) was between 30-50%
Or even the gtx 280 to 480 is around 30-50%.

http://www.tweaktown.com/articles/29...on/index4.html

This was a best case scenario because AMD got to double everything, texture units, ROPS, Shaders.

http://www.anandtech.com/show/2841/17

Drivers might have brought performance up 5% but it also brought up 4890 and 4870 performance as well.

AMD this time is going to at best increase performance by 50% because there is only a 50% increase in real shaders. There is a big but this time around.

AMD is not going to double ROPS and texture units this time around; because of this, AMD will start seeing drops off from that 50%.

Even the wildest optimist on this board beside yourself, doesn't believe an 80% increase will happen because this is the same node.

Much of the purpose of changing to a new architecture again, is trying to make those spec gains linear again. Because if their was only a 1920 shaders difference(using the same technology), their might be a 5% performance increase between the 5870 and 6970, because the 5850 and 5870 perform the same at the same clocks and thus shows, shaders are encountering another bottleneck in the architecture. Changing to a new architecture is going to help the generation pick up some gains again, but not this perfectly linear 80 percent your thinking of. Barts XT performs as well as it is, because it has the ideal configuration(encounters the least amount of bottlenecks) to get maximized performance out of the r600 - r800 architecture. AMD is not going to get this ideal architecture off the bat, as it took AMD 3 and a half years to get there.

BTW, Barts xt is clocked at 900 and not 725mhz.

can't compare HD4890-HD5870.

There was some obvious efficency lost moving to the DX11.. just look at HD5770 vs 5830 vs 4890

There's (logically) only room for improvment in efficency, as we saw with barts.

Still, I agree it will be well below 50% increase in performance if current specs are true..

[tajoh111]

can't compare HD4890-HD5870.

There was some obvious efficency lost moving to the DX11.. just look at HD5770 vs 5830 vs 4890

Yes, I know, but they got to double the specs of everything which makes up for much of the bottlenecks. The 6970 is not going to get that luxury and the biggest increase will be the number of real shaders, hence the size of the chip not reaching gtx 480.

If they can improve the performance by 50% that already shows a huge efficiency jump. I am still thinking about 40% because I think they will keep the chip under 400mm2.

I think the huge increase in efficiency can be easily noticed from the following comparison.

(double TMU, Double ROPS, Double shaders, a tad faster memory) = 44% increase.

50% increase in shaders, significantly faster memory = 40 percent in speed.

Even if they get 40%, considering how they are not beefing up the rest of the architecture as much, it will still be an accomplishment considering they are on the same node.

[Solus Corvus]

Even if they get 40%, considering how they are not beefing up the rest of the architecture as much, it will still be an accomplishment considering they are on the same node.

How do we know what they will be doing to the rest of the architecture?

[tajoh111]

How do we know what they will be doing to the rest of the architecture?

I think AMD still wants to keep the card around 400nm and keep power reasonable. Adding twice as much memory and increase the speed is already going to add alot. 50% more shaders is going to add quite a bit as well. And if these shaders are used properly, they are probably going to consumer more watts pers shader.

[Solus Corvus]

I think AMD still wants to keep the card around 400nm and keep power reasonable. Adding twice as much memory and increase the speed is already going to add alot. 50% more shaders is going to add quite a bit as well. And if these shaders are used properly, they are probably going to consumer more watts pers shader.

By getting rid of one shader in an group that goes mostly unused and spreading the remaining functionality between the other shaders they can increase utilization. That could result in better performance/watt to not have some mostly idle units.

I think we will have to wait and see, there is quite a lot we don't know still.

[LordEC911]

I think AMD still wants to keep the card around 400nm and keep power reasonable. Adding twice as much memory and increase the speed is already going to add alot. 50% more shaders is going to add quite a bit as well. And if these shaders are used properly, they are probably going to consumer more watts pers shader.

They have reworked Cypress' MC and they are a little bit better as far as power efficiency goes. Even though it is double the memory they are double the density, supposedly, and on a lower node which should decrease power consumption slightly compared to the HD5870 2Gb models which I believe added a good ~30-40w over the stock HD5870. Obviously having 2Gb will increase power consumption but it wouldn't be as much as last gen's 2Gb versions.

My figures may be off a bit, but if Cayman XT is for argument's sake 10% slower than a 580, I would expect it to be more than 10% off of the 580's price overall (more like 15-17% less) since they need a bigger incentive to settle for less (same goes for if Cayman XT is faster than the 580, of course...). The top card always carries a bigger premium in respect to price/performance, and I expect amd/nvidia will capitalize on that whichever way the pendulum swings here.

I would expect GTX580 and Cayman to at least be trading blows, since Cayman was targetting +10-15% on a full specced GTX480 w/ 512SPs. So even if they miss the mark a bit or GTX580 gets a bit more performance, there shouldn't be a huge difference in performace. I just hope that they will have Cayman Pro at or slightly above GTX480 levels of performance.

[bill_d]

How do we know what they will be doing to the rest of the architecture?

how do we know anything since the 4870's and the how many sp's war

AMD has just got better and better at keeping secrets

[Lanek]

Yes, I know, but they got to double the specs of everything which makes up for much of the bottlenecks. The 6970 is not going to get that luxury and the biggest increase will be the number of real shaders, hence the size of the chip not reaching gtx 480.

If they can improve the performance by 50% that already shows a huge efficiency jump. I am still thinking about 40% because I think they will keep the chip under 400mm2.

I think the huge increase in efficiency can be easily noticed from the following comparison.

(double TMU, Double ROPS, Double shaders, a tad faster memory) = 44% increase.

50% increase in shaders, significantly faster memory = 40 percent in speed.

Even if they get 40%, considering how they are not beefing up the rest of the architecture as much, it will still be an accomplishment considering they are on the same node.

forget it ... you have no idea how a SP ( or how VLIW work,) and so you have no idea of what a increase of his efficient can be bring or not .... seriously this speak going to nowhere...