AMD Polaris architecture – GCN 4.0

[Mei]

another month of guessing

 

[Seanspeed]

Thats the thing - from what we have seen P10 probably has 2560 shaders at least and if they are running at around 1GHZ with the uarch improvements,I can see why it might be around R9 390/390X level performance.

But if AMD could clock the cores higher,ie,like 1.3GHZ to 1.4GHZ that would be a different kettle of fish. That would be more like GTX980TI level performance. I wonder if that is why we have not seen Polaris released yet??

I cant imagine 300-400Ghz at the top end is going to be the difference between 390 and a 980Ti.

My guess is that 'properly' clocked, P10 should do about 390X level, maybe a bit better. This is where the original 2x performance/watt claims fit in. Conservatively clocked, it might do the newly claimed 2.5x performance/watt, but only perform close to 390 level.

 

[Mtom]

Looking at the Pascal thread Nvidia looks like they will hit 1.9GHZ with their new cards for boost clocks. Its a bit worrying all the leaks we have seen about Polaris seem to be between 800MHZ to 1GHZ or thereabouts.

I wouldn't worry too much.
Its the same as with maxwell..."wooo my 970 can run at 1500MHz!" .....and offers similar performance as a 390 at 1150MHz.

 

[CAT-THE-FIFTH]

I cant imagine 300-400Ghz at the top end is going to be the difference between 390 and a 980Ti.

My guess is that 'properly' clocked, P10 should do about 390X level, maybe a bit better. This is where the original 2x performance/watt claims fit in. Conservatively clocked, it might do the newly claimed 2.5x performance/watt, but only perform close to 390 level.

That would would be 30% to 40% higher clockspeed than an R9 390 or R9 390X,which are around 1GHZ,with the fastest custom cards hitting 1.1GHZ.

Look at the latest TPU review:

http://tpucdn.com/reviews/ASUS/GTX_950/images/perfrel_2560_1440.png

At 2560X1440,the difference between an R9 390X and a GTX980TI is around 30% or thereabouts.

Nvidia has managed to increase the boost clocks massive over AMD on 28NM and AMD has actually managed to regress in some ways.

 

[D.P.]

I cant imagine 300-400Ghz at the top end is going to be the difference between 390 and a 980Ti.

My guess is that 'properly' clocked, P10 should do about 390X level, maybe a bit better. This is where the original 2x performance/watt claims fit in. Conservatively clocked, it might do the newly claimed 2.5x performance/watt, but only perform close to 390 level.

This pretty much. Cutting the clock cuts power dramatically (squared relationship), AMD got a big PR win with the Nano by hand selecting the best chips and cutting the frequency a little to give low power. If they use modest clocks then they will hit their 2x performance per watt marketing. This also leaves space for AMD to release a higher clocked higher performing card.

Chips also are designed to run optimally at certain clock speeds. You can increase the clock speed by making some design decisions. The Pascal GP100 is a very big chip with very high clocks, that is no coincidence. Lower density chips with increased spacing between hot components can allow increased clocking. Looks like nvidia might have done something similar with G104, a slightly bigger chips and facilitating clock speeds.

 

[CAT-THE-FIFTH]

I wouldn't worry too much.
Its the same as with maxwell..."wooo my 970 can run at 1500MHz!" .....and offers similar performance as a 390 at 1150MHz.

The problem is AMD cards are still running at 1GHZ to 1.1GHZ,and remember if the GTX980TI were running at the same clockspeeds as a Fury X it would probably not be as fast.

I really hope AMD gets over the clockspeed wall they have,otherwise they need to keep packing more and more shaders in and hope the APIs can catch up to leverage them.

 

[Seanspeed]

That would would be 30% to 40% higher clockspeed than an R9 390 or R9 390X,which are around 1GHZ,with the fastest custom cards hitting 1.1GHZ.

Look at the latest TPU review:

http://tpucdn.com/reviews/ASUS/GTX_950/images/perfrel_2560_1440.png

At 2560X1440,the difference between an R9 390X and a GTX980TI is around 30% or thereabouts.

Nvidia has managed to increase the boost clocks massive over AMD on 28NM and AMD has actually managed to regress in some ways.

Looks more like a 40% advantage there.

And you dont necessarily get a linear 10% improvement for every 100Mhz clock speed increase.

With a 200Mhz boost to my 970, I'm generally getting about a 10% performance increase.

 

[CAT-THE-FIFTH]

This pretty much. Cutting the clock cuts power dramatically (squared relationship), AMD got a big PR win with the Nano by hand selecting the best chips and cutting the frequency a little to give low power. If they use modest clocks then they will hit their 2x performance per watt marketing. This also leaves space for AMD to release a higher clocked higher performing card.

Chips also are designed to run optimally at certain clock speeds. You can increase the clock speed by making some design decisions. The Pascal GP100 is a very big chip with very high clocks, that is no coincidence. Lower density chips with increased spacing between hot components can allow increased clocking. Looks like nvidia might have done something similar with G104, a slightly bigger chips and facilitating clock speeds.

Looking at tests people have done and what clockspeeds the professional cards run at - the later iterations of GCN,ie,GCN1.1 and 1.2 seem to be optimal around 800MHZ to 900MHZ. Anything after leads to big increases in power consumption from what I gather.

Looks more like a 40% advantage there.

And you dont get a linear 10% improvement for every 100Mhz clock speed increase.

With a 200Mhz boost to my 970, I'm generally getting about a 10% performance increase.

The GTX980TI tested was 26% faster than the R9 390X,and 35% faster than a R9 390. Yes,but at the same time a 30% to 40% increase in core clockspeeds would change things massively.

That is the problem,AMD is having currently.

Nvidia is boosting clockspeeds over the last three generations compared to what AMD is doing and the later AMD cores seem almost maxed out in terms of clockspeed. AMD has slightly regressed. It means they end up plonking more shaders in,which leads to other problems.

Imagine if the Fury X could hit 1.2GHZ to 1.3GHZ,would the GTX980TI look as impressive as it is??

An 800MHZ to 900MHZ clockspeed advantage would be massive for Nvidia.

 

[humbug]

If a Hawaii XT has 2816 Shaders @ 438mm^2 then P10 has 3041 Shaders (+8%)
Samsung 14nm FinFet node is 2x density of TSMC 28nm. P10 measures 232mm^2

So by that measure, and not including any GCN 4.0 architectural performance improvements or performance improvements from faster transistor switching in 14nm FinFet and at the same 1050Mhz clock rates a GTX 980TI would be 22% faster.

Again, only if you ignore improvements from faster transistors and architectural improvements.

 

[AllBodies]

If a Hawaii XT has 2816 Shaders @ 438mm^2 then P10 has 3041 Shaders (+8%)
Samsung 14nm FinFet node is 2x density of TSMC 28nm. P10 measures 232mm^2

So by that measure, and not including any GCN 4.0 architectural performance improvements or performance improvements from faster transistor switching in 14nm FinFet and at the same 1050Mhz clock rates a GTX 980TI would be 22% faster.

Again, only if you ignore improvements from faster transistors and architectural improvements.

According to Samsung, 14LPE was 2x the density, and 14LPP is 'up to' 2.3x.

 

[humbug]

According to Samsung, 14LPE was 2x the density, and 14LPP is 'up to' 2.3x.

Right, LPE is early 14nm, much like 28nm of which GPU's are built on 28nm LPP 14nm are also likely to be built on LPP.

Up scaled that would make P10 about the size of a Fury-X.

 

[AllBodies]

Right, LPE is early 14nm, much like 28nm of which GPU's are built on 28nm LPP 14nm are also likely to be built on LPP.

Up scaled that would make P10 about the size of a Fury-X.

Yeah, if Polaris is the maximum density available on 14LPP, it should be equivalent to a ~530mm2 28nm chip.

Not including arch. improvements or clockspeed improvements from 14nm.

I am looking forward to seeing for sure how Polaris 10 and the GTX 1080 stack up. The theoretical figures are quite tasty

 

[andybird123]

Right, LPE is early 14nm, much like 28nm of which GPU's are built on 28nm LPP 14nm are also likely to be built on LPP.

Up scaled that would make P10 about the size of a Fury-X.

No. GPU's were built on 28nm HP.
28nm LP was available from late 2010 with devices appearing in 2011. 28nm HP became available in late 2011

 

[Overlag]

I wouldn't worry too much.
Its the same as with maxwell..."wooo my 970 can run at 1500MHz!" .....and offers similar performance as a 390 at 1150MHz.

Like the old P4s vs AMD chips. Mhz isnt everything.

 

[tavtavtav]

This pretty much. Cutting the clock cuts power dramatically (squared relationship)

FYI this isn't true. You may have been thinking of voltage, as P=V^2/R.

I believe the frequency - power relationship is linear, which also seems intuitive IMO

 

[D.P.]

FYI this isn't true. You may have been thinking of voltage, as P=V^2/R.

I believe the frequency - power relationship is linear, which also seems intuitive IMO

It is, but if you cut the frequency then you can additionally cut the voltage so you get the benefit of both, which is O (n^2).

 

[GFX-Kid]

AMD need to leak something to keep us excited until official word is out

If nvidia actually launch shortly, AMD may lose a lot of sales again. I for one may pick up a pascal as long as nvidia do not take the **** too much with the pricing.

But if I know I can get similar or better performance from amd for cheaper, I will certainly hold out

 

[Gregster]

Such a shame AMD were not 6 months ahead and I would have liked to have seen them out first but if they are late, not good at all.

 

[GFX-Kid]

Such a shame AMD were not 6 months ahead and I would have liked to have seen them out first but if they are late, not good at all.

I don't know about the 6 month bit, but certainly agree they should not be launching behind nvidia. If they have anything worth leaking, they should do so now...

 

[Relentless81]

I'm rapidly losing interest, they need to release some solid info because there are other things I can buy, my 290X can stay in my PC for the rest of summer

All the rumors and speculation have grown tedious for me, it would be nice to know what exactly is being released so I can make a decision, especially seeing as though its supposed to be next month when they are released