AMD Polaris architecture – GCN 4.0

[Baboonanza]

Have any of the Polaris cards been physically pictured yet? Wondering about physical size for my mini itx...

I want the same info. I would guess that if they are using gddr they will not be much smaller than 3xx cards. Possibly slightly shorter due to having a lower power draw? 25cm+ is likely IMO.

For a nice small card well need something with hbm/hbm2. And asfaik nobody knows if there will be one before vega.

 

[Davedree]

I want the same info. I would guess that if they are using gddr they will not be much smaller than 3xx cards. Possibly slightly shorter due to having a lower power draw? 25cm+ is likely IMO.

For a nice small card well need something with hbm/hbm2. And asfaik nobody knows if there will be one before vega.

Sapphire make an itx 2gb 285/4gb 380 so it can be done. With Amd's Nano i'd like to think that AMd will look into itx as a reference design for the polaris.
I currently use an msi r7 360 in a htpc and polaris itx is of interest to me.

 

[CAT-THE-FIFTH]

Emm?? You do realise the 212mm2 Pitcairn chip in the HD7870 was 10% faster than an HD6970 and consumed 80% less power on average?? Pitcairn was nearly half the size of the chip in the HD6970 and that was going from 40NM to 28NM AND IT HAD MORE TRANSISTORS TOO. This is going from 28NM to second generation 14NM wih Finfets.

The 14NM process has double the transistor density with a 60% to 70% reduction in power.

So,that means if you shrunk Tonga directly down with no uarch changes,you could have an R9 380X probably consuming no more than 100W. That would put it close to 180MM2.

Tonga is a 360MM2 chip,but even the R9 380X does not have the 384 bit memory controller enabled.The R9 380 has 2/3 of the memory controller enabled and 87.5% of the shaders enabled.

Polaris 11 is also using a 128 bit memory controller too,so less RAM chips and less power consumption over a partially disabled 384 bit one.

It also lots of other changes which have been detailed - one of them appears to be power gated CPUs,etc.

So taking that into consideration Polaris 11,can probably easily match a GTX960 or R9 380 and have massively lower power draw.

If you even look at the average power figures for the R9 380X for example:

url] A 60% reduction in that would...ill be different Polaris 11 based cards too??

 

[andybird123]

You're not talking apples for apples CAT
Bru was talking about TDP, the TDP for the 380 is 190W, you're using gaming/average usage which is not the same thing, even the charts you linked to show that other test methods have that same card at higher rates than you're cherry picking

 

[CAT-THE-FIFTH]

You're not talking apples for apples CAT
Bru was talking about TDP, the TDP for the 380 is 190W, you're using gaming/average usage which is not the same thing, even the charts you linked to show that other test methods have that same card at higher rates than you're cherry picking

What cherry picking - you and him are cherry picking. That is card power consumption from TPU who tends to be more towards Nvidia.

You are both on purpose confusing it,and trying to confuse TDP which is a cooling specification with actual board power and actual power consumption,which is what AMD was showing with Polaris 11,with the balanced demo. I don't see any of you both doing that in the Pascal thread too to thrash the thread.

This is what you don't seem to understand and cherry-ignore,the previous generations and you are on purpose ignoring what we saw from 40NM to 28NM - the HD7870 had nearly twice the performance per watt of the HD6970 with similar performance and similar amount of transistors.

Both of you are cherry-confused,that it is the transistor budget which is the most important thing here,and seemed to be so cherry-confused that you have seemed to forgotten the last decade of node shrinks. Its not DIE SIZE but transistor budget,which is the most important thing here.

You are on purpose cherry-confusing things,when you very well know what a direct shrink of even Tonga to 14nm Finfet would bring and instead want to back up your mate,who keeps repeating the same thing(even though he does not seem to understand what he is saying) and even cherry-ignores cards like the R7 360E which don't fit what he said(which are based on Bonaire).

Yet,in the Pascal thread,I don't see any of the false cherry-confusing happening by both of you,and we know why.

But I also like to know why both of you want to cherry pick Nvidia only doing something like Maxwell with its huge performance/watt improvements on the same node,but this cannot be done by AMD on a new node with Polaris.

Seems cherry-convenient.

 

[r7slayer]

What cherry picking - you and him are cherry picking. That is card power consumption from TPU who tends to be more towards Nvidia.

You are both on purpose confusing it,and trying to confuse TDP which is a cooling specification with actual board power and actual power consumption,which is what AMD was showing with Polaris 11,with the balanced demo. I don't see any of you both doing that in the Pascal thread too to thrash the thread.

This is what you don't seem to understand and cherry-ignore,the previous generations and you are on purpose ignoring what we saw from 40NM to 28NM - the HD7870 had nearly twice the performance per watt of the HD6970 with similar performance and similar amount of transistors.

Both of you are cherry-confused,that it is the transistor budget which is the most important thing here,and seemed to be so cherry-confused that you have seemed to forgotten the last decade of node shrinks. Its not DIE SIZE but transistor budget.

You are on purpose cherry-confusing things,when you very well know what a direct shrink of even Tonga to 14nm Finfet would bring and instead want to back up your mate,who keeps repeating the same thing(even though he does not seem to understand what he is saying) and even cherry-ignores cards like the R7 360E which don't fit what he said(which are based on Bonaire).

Yet,in the Pascal thread,I don't see any of the false cherry-confusing happening by both of you,and we know why.

But I also like to know why both of you want to cherry pick Nvidia only doing something like Maxwell with its huge performance/watt improvements on the same node,but this cannot be done by AMD on a new node with Polaris.

Seems cherry-convenient.

You seem cherry-mad bro!

Sorry couldn't resist
However i agree with what your saying.

 

[CAT-THE-FIFTH]

You seem cherry-mad bro!

Sorry couldn't resist
However i agree with what your saying.

Tis true!!

The only way I can't see Polaris 11 in its MAXED out form not hitting GTX960/R9 380 level performance is if its much smaller than the estimates say it is - like 100MM2,and not 125MM2 to 150MM2. Transistor budget will be the determining factor here.

Even if Polaris 11 were exactly half a 232MM2 Polaris 10 at 116MM2 and Polaris 10 was basically R9 390X level performance,that would place Polaris 11 at GTX950 level performance,if the scaling were exact(most of the GCN1.0 cards had this apart from Tahiti),and even then we don't know where the optimal performance level is for Polaris,so it could be higher(it is only a 20% difference between a GTX950 and a GTX960 for example).

The other thing is too - Polaris 10 on 14NM LPP is probably having a similar transistor budget(or maybe a tad higher) when compared to Hawaii.

So,if they are targetting R9 390X level performance,going from the last rumour on a Chinese forum,that would be coming with a large improvement in performance/watt. No different than the HD6970 to the HD7870 which Polaris 10 looks like it is more like.

This is why I also think Nvidia will be having the higher performing options until Vega is out. They are throwing more transistors at the problem than AMD.

 

[CAT-THE-FIFTH]

Also,to AB and Bru,I am not going to argue with you about this anymore and if you still disagree,we will have to agree to disagree and leave it at that. People can read what is posted and make their own minds up.

 

[sprite-]

I read this thread on rare occasions.

Have we got any realistic release date, models, prices or benchmarks yet? Or are people still just spewing internet vomit on the subject? Want to upgrade my 290x.

 

[Seanspeed]

Tis true!!

The only way I can't see Polaris 11 in its MAXED out form not hitting GTX960/R9 380 level performance is if its much smaller than the estimates say it is - like 100MM2,and not 125MM2 to 150MM2.

I agree. Though we should keep in mind we've only once heard something about it being 150mm, while before the rumors were saying it was smaller. So I dont think it's safe to say it's *likely* Polaris 11 will hit those targets you're mentioning, but it's definitely a possibility.

I read this thread on rare occasions.

Have we got any realistic release date, models, prices or benchmarks yet? Or are people still just spewing internet vomit on the subject? Want to upgrade my 290x.

Not many concrete facts at the moment on actual products, but the speculation being discussed here is far from 'internet vomit' with no substance for the most part. There's enough to work with to form a rough idea of what we could be getting.

As for you looking to upgrade your 290X, Polaris may not be what you're looking for if you want a significant upgrade. It doesn't seem to be aiming at the higher end segment. But we'll see.

 

[flopper]

I read this thread on rare occasions.

Have we got any realistic release date, models, prices or benchmarks yet? Or are people still just spewing internet vomit on the subject? Want to upgrade my 290x.

dx12 hitman 60fps Polaris 232mm2 core die.
Thats about it as far.
Thats in the furyx ballpark btw.

 

[Seanspeed]

Thats in the furyx ballpark btw.

It's more like 390 ballpark, actually.

And that's assuming max settings were used in the Polaris demo.

 

[bru]

So the NVidia 980Ti is not a 250W part as it states on NVidia's own website but it is only a 211W part because that is what it pulls in average gaming.

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-980-ti/specifications

These figures are the industry standard way of being measured, not what these cards use in average gaming, that is how the system works.

You seem to think that the density improvements come with ZERO power consumption improvements.

Err no they come with a 2x per/W


As for all my ramblings, the same as anything else I could be miles off, AMD might do something really bizarre and make each chip into the shape of a banana for all I know, it is all speculation and rumour. At least we don't have too long to wait now till we finally get some answers.

 

[FoxEye]

I read this thread on rare occasions.

Have we got any realistic release date, models, prices or benchmarks yet? Or are people still just spewing internet vomit on the subject? Want to upgrade my 290x.

Absolutely nothing. Which means release must still be months away, because you can't keep such a tight lid on info if cards have been sent to distributors, etc.

I'm not betting on a June release at this point

 

[CAT-THE-FIFTH]

I agree. Though we should keep in mind we've only once heard something about it being 150mm, while before the rumors were saying it was smaller. So I dont think it's safe to say it's *likely* Polaris 11 will hit those targets you're mentioning, but it's definitely a possibility.

It comes from the Videocardz editor - he did some extra measurements and mentioned it in the comments to the new article he did,which confirmed the 232MM2 rumour.

That would make P11 64% of a Polaris 10 if 150MM2. On one of the Chinese forums,known for reasonably accurate leaks(still could be wrong),they said P10 was 97% of a R9 390X. That would place it at roughly R9 380 performance or GTX960 level performace(slight underestimate on my part too).

If you look at the original GCN1.0 launch Cape Verde was 58% of the size of Pitcairn. The HD7770 was 60% of the performance of the HD7870.

If it is 120MM2 it would scale to around GTX950 level.

Then you need to consider that both 14NM and 16NM are not only a one node shrink but also have Finfets which help towards performance/watt improvements AND AMD and TSMC are using second generation versions too. So in some ways,its probably not a normal shrink.

Considering AMD only has started trialing some of the Maxwell level power consumption reduction features in Fiji and Polaris is meant to have things like powergated CUs,etc I can see them have better than normal jumps in performance/watt by their standards.

You also need to consider AMD massively refined its firmware too. The bus powered R7 360 is a slightly cut down R7 260X,but with the latest PowerTune,it hits GTX750 level performance and power consumption.

That would make the bus powered R7 360 the highest performance/watt AMD graphics card of this generation followed by the Nano.

You also need to consider this - Nvidia has made a bus powered GTX950. I am surprised if AMD cannot beat that convincingly with a node shrink and a new uarch.

 

[Seanspeed]

It comes from the Videocardz editor - he did some extra measurements and mentioned it in the comments to the new article he did,which confirmed the 232MM2 rumour.

In the article, it sounded like he was acknowledging quite clearly that he was having to make unsafe assumptions in order to make the measurements and so to not take them as gospel.

 

[CAT-THE-FIFTH]

Yea, but if you read the article, he was very cautious in saying the measurements were anything close to confirmed as he was making a bunch of assumptions.

Going from past AMD GPUs at launch the lower end jobbie is between 50% to 60% of the one above it(last two node shrinks). Even at half Polaris 10 size,it should have no issue hitting GTX950 level performance and that is being conservative. Even an R9 270X does that,and that is just over 200MM2 on 28NM with a 256 bit memory controller. Shrinking down AMD Pitcairn on 14NM,would yield a 100MM2 GPU and Pitcairn has much higher performance per watt than Tonga which regressed. Me choosing Tonga was done on purpose as it is the worst case scenario in terms of performance/mm2 and performance/watt.

Using AMD Pitcairn as a baseline would make Polaris 11 look massively better in all metrics.

GTX950 level performance is a very conservative estimate.

Now,look at the other end of the spectrum - AMD Pitcairn which already has GTX950 level performance.

60% less power consumption,would mean an R7 270X would be consuming under 60W:

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

Look at that chart - it is the bus powered GTX950,Nvidia recently produced which consumes 75W,ie,the PCI-E slot limit.

Are people really believing,that Polaris 11,with a new uarch,and a 14NM LPP process can barely get 20% better performance/watt than a GTX950??

Because that is what a bus powered GTX960/R9 380 level Polaris 11 would be doing as a GTX960 is 20% faster than a GTX950 and I have a GTX960 myself.

If that is all what AMD can do with a new uarch and process node,I would be worried what happens when Nvidia launches the GP106 and lower end chips in the Pascal range.

 

[Jono8]

Are people really believing,that Polaris 11,with a new uarch,and a 14NM LPP process can barely get 20% better performance/watt than a GTX950??

Who knows, but you have to remember AMD are playing catch up big time in terms of perf per watt compared with Maxwell. Only the Fury range could get close at the high end, and that was only because they used 4gb of HBM (compared with 6gb of DDR5 in the 980Ti).

Everywhere else , Maxwell was hugely more efficient.

 

[Seanspeed]

Going from past AMD GPUs at launch the lower end jobbie is between 50% to 60% of the one above it(last two node shrinks). Even at half Polaris 10 size,it should have no issue hitting GTX950 level performance and that is being conservative. Even an R9 270X does that,and that is just over 200MM2 on 28NM with a 256 bit memory controller. Shrinking down AMD Pitcairn on 14NM,would yield a 100MM2 GPU and Pitcairn has much higher performance per watt than Tonga which regressed. Me choosing Tonga was done on purpose as it is the worst case scenario in terms of performance/mm2 and performance/watt.

Using AMD Pitcairn as a baseline would make Polaris 11 look massively better in all metrics.

GTX950 level performance is a very conservative estimate.

Now,look at the other end of the spectrum - AMD Pitcairn which already has GTX950 level performance.

60% less power consumption,would mean an R7 270X would be consuming under 60W:

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

Look at that chart - it is the bus powered GTX950,Nvidia recently produced which consumes 75W,ie,the PCI-E slot limit.

Are people really believing,that Polaris 11,with a new uarch,and a 14NM LPP process can barely get 20% better performance/watt than a GTX950??

Because that is what a bus powered GTX960/R9 380 level Polaris 11 would be doing as a GTX960 is 20% faster than a GTX950 and I have a GTX960 myself.

If that is all what AMD can do with a new uarch and process node,I would be worried what happens when Nvidia launches the GP106 and lower end chips in the Pascal range.

Dude, I'm not arguing with you about what the potential performance levels would be if 'blah blah blah'. I think your figures are generally pretty spot on with what I think as well.

Just saying that what size Polaris 11 will actually be, which as you say - will dictate what the resulting performance is - is definitely still up in the air.

 

[Troezar]

7970CF Replacement

First off this is speculation, not unusual for this section of the forum! I've been quite happy with my cards for the last 4+ years but with new ones on the horizon and DX12 here I'm thinking of replacing them and going for a single card.

Any chance that I will be able to get better performance for £300 with the next gen? I currently game at 1440p and 60Hz with my overclocked i5 2500k.