AMD's next GPU uarch is called "Polaris"

[3DVagabond]

Well if you compare Hawaii vs Fiji at 1080p and then at 4K you will see that Fiji is scaling better at 4K than Hawaii and they both have the same 64 ROPs.

So Fiji cannot be ROP starved at 1080p. Perhaps its the API overhead as you say at 1080p but certainly not the ROPs.

I agree. (In my not so expert opinion. )

 

[desprado]

Ehm, Maxwell is what currently on the market TODAY, that is what they can compare with. Lets see NVIDIA compare Pascal to GTX950 at the same game same settings and Compare then

If that is the case than GCN 4.0 is already beaten in efficiency by Pascal.

 

[AtenRa]

If that is the case than GCN 4.0 is already beaten in efficiency by Pascal.

Without even seeing an NVIDIA Pascal Gaming performance/watt slide you come to this conclusion ??? fine, I on the other hand will wait for both to come out and then make a conclusion about that.

 

[3DVagabond]

The problem is that nobody buys reference 980 Ti cards, and every single factory OC'd card stomps the Fury X. If AMD isn't making any major changes this time, then we'll still be in a situation where Nvidia cards have 10-15% factory overclock with another 15% or more of headroom while AMD will continue to top out at 15% MAX overclock after adding a ridiculous amount of voltage. Nothing will change.

We're talking a completely new GPU with a new uarch and a new process. We have no idea how it will clock. Again, Maxwell vs Fiji is irrelevant. It has no place in this discussion.

 

[Techhog]

AMD PR is in such a desperation that they did not waited for pascal to see its performance per watt.

Lol AMD made comparison of Maxwell and GCN 4.0 in performance per watt.

Pascal is 2X more efficient than Maxwell + advantage of new node.

First of all, are you saying that they should just avoid talking about anything until Pascal launches? If so, that's just stupid. You use what you have.

Second, I don't know why some of you think that 2x improvement is from the architecture alone. Seriously, that would mean close to a 4x improvement in efficiency. If that happens, I'll buy a card for everyone here! Please! Nvidia's engineers aren't so great that they can defy the laws of physics. I'd actually argue that they're worse than AMD's in some ways, considering the fact that AMD was able to compete so well even with a mostly aging architecture.

 

[desprado]

Without even seeing an NVIDIA Pascal Gaming performance/watt slide you come to this conclusion ??? fine, I on the other hand will wait for both to come out and then make a conclusion about that.

Nvidia said 3 or 4 times that Pascal is 2X efficient in performance per watt than Maxwell without new node. However, even if Maxwell is used in new node and gets a little bump in efficiency than it can beat GCN 4.0 in efficiency.

 

[lixlax]

Do they even have the new silicon in hand or is this data based on simulations?

 

[AtenRa]

Also, after those new slides were published, it seams one of the new dies will be very small. Something like 100mm2 ??

So, what about the second die ??? will it be close to 200mm2 or even higher ???

 

[desprado]

First of all, are you saying that they should just avoid talking about anything until Pascal launches? If so, that's just stupid. You use what you have.

Second, I don't know why some of you think that 2x improvement is from the architecture alone. Seriously, that would mean close to a 4x improvement in efficiency. If that happens, I'll buy a card for everyone here! Please! Nvidia's engineers aren't so great that they can defy the laws of physics. I'd actually argue that they're worse than AMD's in some ways, considering the fact that AMD was able to compete so well even with a mostly aging architecture.

Have you seen Pascal being compared with GCN 4.0 by Nvidia?

Nvidia is comparing Pascal with Maxwell and that is what AMD should do.Rather than making GCN 4.0 look more weaker against pascal.

 

[AtenRa]

Do they even have the new silicon in hand or is this data based on simulations?

According to this slide they seem to have actual hardware on hand.

 

[littleg]

For a home user / gamer I would think Perf/$ would be the most important metric anyway. If the real world power differences between the competing cards are that much of an issue for you then you're probably better off saving your money to use on something else.

 

[dogen1]

Actually it's not, because if ROPs were the bottleneck, the performance will drop off as high resolution is used as well as MSAA or DSR/VSR.

MSAA maybe, but higher resolutions require more ALU/TMU resources as well as ROP.

We're seeing the opposite, where performance is poor at 1080p and improves at higher resolutions.

The bottleneck with Fury X is shader uptime, the front-end designed for Hawaii/Tonga is made to drive 4K shaders, way too many for it to feed.

Is that true though?

http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4

Tests show it can hit it's full 8.6TFlops, so that can't be right.(Unless I'm missing something here..)

 

[AtenRa]

Nvidia said 3 or 4 times that Pascal is 2X efficient in performance per watt than Maxwell without new node. However, even if Maxwell is used in new node and gets a little bump in efficiency than it can beat GCN 4.0 in efficiency.

Yea that could be attributed to DX-12 vs Maxwell

 

[ShintaiDK]

14nm and tweaks?

GCN 1.3 as it would be called in the common tongue. R&D mainly went to the CPU division.

Also, after those new slides were published, it seams one of the new dies will be very small. Something like 100mm2 ??

So, what about the second die ??? will it be close to 200mm2 or even higher ???

It starts to show what is expected, consumer parts will be based on very small dies due to cost structure. At least we get great performance/watt, maybe not so much performance.

 

[desprado]

GCN 1.3 as it would be called in the common tongue. R&D mainly went to the CPU division.



It starts to show what is expected, consumer parts will be based on very small dies due to cost structure. At least we get great performance/watt, maybe not so much performance.

And they have very few investor and DGPU customer which has also effected their R&D.

 

[AtenRa]

And they have very few investor and DGPU customer which has also effected their R&D.

So if NVIDIA will make small dies it will be attributed to low R&D as well ??

 

[Techhog]

Have you seen Pascal being compared with GCN 4.0 by Nvidia?

Nvidia is comparing Pascal with Maxwell and that is what AMD should do.Rather than making GCN 4.0 look more weaker against pascal.

... What? Are you disagreeing with yourself now?

 

[desprado]

... What? Are you disagreeing with yourself now?

What i mean to say is that they need to compare it with last GCN models rather then with Maxwell.

If AMD thinks that GCN 4.0 competition is Maxwell than they are doing it totally right. However, if it is a competitor of Pascal then they are only making it look weaker.

 

[PhonakV30]

Nvidia said 3 or 4 times that Pascal is 2X efficient in performance per watt than Maxwell without new node. However, even if Maxwell is used in new node and gets a little bump in efficiency than it can beat GCN 4.0 in efficiency.

If 2x efficient in performance per watt , then Pascal doesn't have Fully hardware Async Compute.Hardware schedulers consume more power.If Pascal become like GCN then "efficient in performance per watt" is lower than 2x or maybe 1.5x.

 

[desprado]

If 2x efficient in performance per watt , then Pascal doesn't have Fully hardware Async Compute.Hardware schedulers consume more power.If Pascal become like GCN then "efficient in performance per watt" is lower than 2x or maybe 1.5x.

you know better than nvdia and AMD even better than both of their CEO and employees.

 

[dacostafilipe]

What i mean to say is that they need to compare it with last GCN models rather then with Maxwell.

If AMD thinks that GCN 4.0 competition is Maxwell than they are doing it totally right. However, if it is a competitor of Pascal then they are only making it look weaker.

As far as I know, we don't have Pascal hardware available to compare to ... or did I miss something? :sneaky:

 

[PhonakV30]

you know better than nvdia and AMD even better than both of their CEO and employees.

atm GCN is only GPU uarch that is Fully compatible with DX12 and excellent operation in VR.Can pascal handle Parallel workloads like GCN ? they have to add ton of features like Latest data latch from AMD , fine-grained preemption , Resource Binding Tier 3 , etc.. , adding more features = More transistors

I know that Pascal won't be 10X faster than Maxwell but nvidia insists.Do i believe them or Do I know better than them ?

http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

 

[Glo.]

Does that mean, that we will have GTX970/R9 390 performance packed in 86W?

If so, that is around 70% decrease in power consumption, compared to R9 390.

 

[AtenRa]

Does that mean, that we will have GTX970/R9 390 performance packed in 86W?

If so, that is around 70% decrease in power consumption, compared to R9 390.

I believe that is close enough.

 

[desprado]

atm GCN is only GPU uarch that is Fully compatible with DX12 and excellent operation in VR.Can pascal handle Parallel workloads like GCN ? they have to add ton of features like Latest data latch from AMD , fine-grained preemption , Resource Binding Tier 3 , etc.. , adding more features = More transistors

I know that Pascal won't be 10X faster than Maxwell but nvidia insists.Do i believe them or Do I know better than them ?

http://wccftech.com/nvidia-pascal-gpu-gtc-2015/

Pascal will never be 10X of Maxwell in terms of performance and efficiency and Nvidia never said that. Pascal is only 10X in term of deep learning, data etc.