Polaris: 2x Perf/Watt of Current AMD mainstream GPUs

[Glo.]

again i will just leave this here............

do you see what amd means ?

So that would mean we are looking at around 55GFLOPs/watt on new gen. of GPUs. 2560 GCN core GPU with 1.266 GHz gives 6.5 TFLOPs, and at 125W of power consumption it gives 52 GFLOPs/watt.

 

[sirmo]

I do not, as the image doesn't show.

I missed it too, but I think this was it.

 

[Glo.]

I think I have cracked it. 2560 GCN cores with 1450 MHz core clock. 7,4 Tflops od compute power, with 125W od TDP and 59 GFLOPs\watt. Now everything we know makes sense. Faster than some of previous gen GPUs, with room for improvement with Vega, both on perf\watt, core clock, power consption and core amount. Fast enough to be sold for 399$. And deffinitely faster than GTX 1070. It might be also very close to 1080...
And for sure most power efficient GPU there is.

 

[Arachnotronic]

I think I have cracked it. 2560 GCN cores with 1450 MHz core clock. 7,4 Tflops od compute power, with 125W od TDP and 59 GFLOPs\watt. Now everything we know makes sense. Faster than some of previous gen GPUs, with room for improvement with Vega, both on perf\watt, core clock, power consption and core amount. Fast enough to be sold for 399$. And deffinitely faster than GTX 1070. It might be also very close to 1080...
And for sure most power efficient GPU there is.

I think this will be proven incorrect.

 

[Glo.]

I think this will be proven incorrect.

Fair enough, but at least you could have given more specifics and reasons: why...

 

[maddie]

I think I have cracked it. 2560 GCN cores with 1450 MHz core clock. 7,4 Tflops od compute power, with 125W od TDP and 59 GFLOPs\watt. Now everything we know makes sense. Faster than some of previous gen GPUs, with room for improvement with Vega, both on perf\watt, core clock, power consption and core amount. Fast enough to be sold for 399$. And deffinitely faster than GTX 1070. It might be also very close to 1080...
And for sure most power efficient GPU there is.

I am one who expect close to or even equal to a 1070, and have argued for this, but we don't KNOW anything with precision.

You are building a case with a detailed conclusion without detailed knowledge. The best anyone can do with the limited rough information is to give a range of probable outcomes and assign probabilities. To do otherwise is to deceive oneself.

We have the following. 2X & 2.5X perf/W increase. Highest increase in performance for AMD ever. the most power efficient GPUs, 232mm^2 die size, 2048 or 2304 or 2560 shaders.

I don't know about you, but for myself, assigning an accurate value to these is pure fantasy.

 

[swilli89]

I am one who expect close to or even equal to a 1070, and have argued for this, but we don't KNOW anything with precision.

You are building a case with a detailed conclusion without detailed knowledge. The best anyone can do with the limited rough information is to give a range of probable outcomes and assign probabilities. To do otherwise is to deceive oneself.

We have the following. 2X & 2.5X perf/W increase. Highest increase in performance for AMD ever. the most power efficient GPUs, 232mm^2 die size, 2048 or 2304 or 2560 shaders.

I don't know about you, but for myself, assigning an accurate value to these is pure fantasy.

Yeah its been your typical pre-release speculation war as has been the case for what seems like decades now. Add to the fact everyone has a case of blue balls from being stuck on 28nm for so long.. Fighting assumed performance vs assumed performance, round and round it goes.

 

[Arachnotronic]

Fair enough, but at least you could have given more specifics and reasons: why...

Just comes off as wishful thinking, especially the superlatives (i.e. Most efficient GPU out there).

Anyway, Polaris updates coming soon so hopefully we'll know.

 

[Glo.]

Just comes off as wishful thinking, especially the superlatives (i.e. Most efficient GPU out there).

Anyway, Polaris updates coming soon so hopefully we'll know.

Erm... There is no GPU than would give more than 45 GFLOPs/watt currently, so 59GFLOPs/watt GPU would be most power efficient...

 

[maddie]

Erm... There is no GPU than would give more than 45 GFLOPs/watt currently, so 59GFLOPs/watt GPU would be most power efficient...

The 59 GFLOPs/W is your estimate. Why do you think you can't be mistaken? Seriously.

 

[Glo.]

The 59 GFLOPs/W is your estimate. Why do you think you can't be mistaken? Seriously.

I can be. Why do you think I do not consider this incorrect?

 

[maddie]

I can be. Why do you think I do not consider this incorrect?

When I asked "Why do you think you can't be mistaken?", I meant for you to give some sort of reasoning that led you to that conclusion.

What is your starting assumptions and reasoning for arriving at 59GFLOPs/W?

 

[Qwertilot]

Because you, well posted then defended it? I'm confused 🙂
(Changing your mind is of course definitely allowed!).

 

[Glo.]

When I asked "Why do you think you can't be mistaken?", I meant for you to give some sort of reasoning that led you to that conclusion.

What is your starting assumptions and reasoning for arriving at 59GFLOPs/W?

Erm... 2560 GCN cores, with 1.45 GHz give 7.4 TFLOPs of compute power. Divide this number by 125W and you get 59 GFLOPs/Watt.

All was in that post above, you quoted. This is of course absolutely theoretical, based on what AMD touted, and with getting some common sense, for this calculations.

Guys, I can be wrong. I do accept and understand that.

 

[maddie]

Erm... 2560 GCN cores, with 1.45 GHz give 7.4 TFLOPs of compute power. Divide this number by 125W and you get 59 GFLOPs/Watt.

All was in that post above, you quoted. This is of course absolutely theoretical, based on what AMD touted, and with getting some common sense, for this calculations.

Guys, I can be wrong. I do accept and understand that.

OK.

3 numbers are the core assumptions. 2560 shaders, 1.45 GHz and 125W.

These can't be stated as proven. IF they are correct, then yes, you will get 59GFLOPs/W, but that is still a big IF, therefore speculation. In any case, AFAIK, AMD never made those claims, or any at all, relating to shader count, frequency and TDP.

 

[Glo.]

OK.

3 numbers are the core assumptions. 2560 shaders, 1.45 GHz and 125W.

These can't be stated as proven. IF they are correct, then yes, you will get 59GFLOPs/W, but that is still a big IF, therefore speculation. In any case, AFAIK, AMD never made those claims, or any at all, relating to shader count, frequency and TDP.

And where did I stated that it as anything but speculation?

 

[raghu78]

I think the 2x perf/watt number touted by AMD management refers to theoretical FLOPS while Raja Koduri's 2.5x perf/watt refers to real world performance in games. We know that AMD's problem with existing cards is their theoretical FLOPS are not well translated to real world performance in games due to bottlenecks and inefficiency(poor shader utilization). Thats why Maxwell was able to beat GCN with vastly lower FLOPS but better efficiency. AMD's key goal with Polaris is to translate the FLOPs into the highest real world performance in the most efficient manner. I think thats exactly what we are likely to see.

 

[maddie]

I think the 2x perf/watt number touted by AMD management refers to theoretical FLOPS while Raja Koduri's 2.5x perf/watt refers to real world performance in games. We know that AMD's problem with existing cards is their theoretical FLOPS are not well translated to real world performance in games due to bottlenecks and inefficiency(poor shader utilization). Thats why Maxwell was able to beat GCN with vastly lower FLOPS but better efficiency. AMD's key goal with Polaris is to translate the FLOPs into the highest real world performance in the most efficient manner. I think thats exactly what we are likely to see.

Agreed.

The new command processor will hopefully feed those cores more efficiently. Fiji is a good example of what happens if you don't. Also if they truly have fine grained power control, it must be an enormous task to optimize those control algorithms. The possible permutations.

Koduri did hint to this during the first P11 showing, when he stated that many power optimizations still needed to be completed.

 

[C@mM!]

I saw this incredible shitpost based on a tonne of guesswork a little while ago, but may be indicative as it sits mostly in line with known performance.

From what I can see, it normalised die size based on fab, and worked out clockspeed based on TDP of the 1080, then normalised for a TDP of 150w (based off Nvidia's 16nm node).

By the looks, that'd put polaris 10 in 1070 territory.

 

[maddie]

I saw this incredible shitpost based on a tonne of guesswork a little while ago, but may be indicative as it sits mostly in line with known performance.

From what I can see, it normalised die size based on fab, and worked out clockspeed based on TDP of the 1080, then normalised for a TDP of 150w (based off Nvidia's 16nm node).

By the looks, that'd put polaris 10 in 1070 territory.

Someone with a lot of imagination at work here.

390X die size should be 438 not 359.
What is normalized die size? Normalized to what?
Mhz by nm. ??????????

I'm lost, where's Alice?

 

[casiofx]

Someone with a lot of imagination at work here.

390X die size should be 438 not 359.
What is normalized die size? Normalized to what?
Mhz by nm. ??????????

I'm lost, where's Alice?

From what I can tell, the normalized die size means to convert the 16nm and 14nm dies to 28nm equivalent sizes...

The mhz by nm is mystery, but total performance looked like mhz times die size... Ahaha