AMD Polaris architecture – GCN 4.0

[fs123]

It would be suicide to launch a mid Polaris gpu that's the same or barely faster than a FuryX. AMD need to release a card that is at least 40% faster than a FuryX to stay in the running.

 

[ICDP]

It would be suicide to launch a mid Polaris gpu that's the same or barely faster than a FuryX. AMD need to release a card that is at least 40% faster than a FuryX to stay in the running.

Pretty much how it has happened with each new GPU launched on a new node since the dawn of GPUs. But somehow we are to believe Foxeye's twisted logic that AMD plans to release only low and mid-range Polaris GPUs. Despite the fact that tech journalists have confirmed they were shown the new "high-end enthusiast" level GPU behind closed doors.

The mind boggles sometimes at the lack of applied logic and fact twisting an individual will go to in order to make facts fit their personal agenda.

 

[Gregster]

It would be suicide to launch a mid Polaris gpu that's the same or barely faster than a FuryX. AMD need to release a card that is at least 40% faster than a FuryX to stay in the running.

I agree and the same for Nvidia. Nobody wants to buy the same as what we already have and we have been milked long enough on 28nm and time for a proper jump in performance and I don't mind paying but not getting trapped in the spoon fed increments we have had on 28nm.

 

[humbug]

I agree and the same for Nvidia. Nobody wants to buy the same as what we already have and we have been milked long enough on 28nm and time for a proper jump in performance and I don't mind paying but not getting trapped in the spoon fed increments we have had on 28nm.

+1, lets have some real upgrades.

 

[ICDP]

+1, lets have some real upgrades.

My expectation is for around 30% over Fury X and stock 980Ti but with decent OC headroom taking it to ~45% faster. I fully anticipate a lot of the extra headroom will go towards reducing TDP which will give fairly decent OC potential.

 

[humbug]

My expectation is for around 30% over Fury X and stock 980Ti but with decent OC headroom taking it to ~45% faster. I fully anticipate a lot of the extra headroom will go towards reducing TDP which will give fairly decent OC potential.

For the Enthusiast end?

If i can get +40% or more over my 970 for about the same money i'm happy.

We deserve that at least, as Greg said we have put up with more than enough increments of pretty much the same for too long.

 

[Scougar]

I think they did a good job stretching it out so long at the same level. Frankly, I don't care that much about a process number, just the performance, features, price etc. If either company released a card that was 40% faster for the same price, would anyone be that weird that they wouldn't appreciate it?

This thread has seen more junk talk about what console means than anything useful.

 

[ICDP]

For the Enthusiast end?

If i can get +40% or more over my 970 for about the same money i'm happy.

We deserve that at least, as Greg said we have put up with more than enough increments of pretty much the same for too long.

I would say ~30% increase across all price ranges. So at the same price point expect 30% stock to ~ 45% with overclocking. This would put the 970 or 390 equivalent at Fury X/980Ti or marginally better speeds. The Fury X replacement would be ~60%+ faster than 970.

 

[AlamoX]

oil is about to get as cheap as barrel of water, the petro dollar is having seizures, soon we all will buy cards for couple million dollars xD, how is that for cheap
but i seriously doubt the new polaris or pascal would be cheaper than current GPU equivalent.

 

[ICDP]

oil is about to get as cheap as barrel of water, the petro dollar is having seizures, soon we all will buy cards for couple million dollars xD, how is that for cheap
but i seriously doubt the new polaris or pascal would be cheaper than current GPU equivalent.

Agreed. Unless Nvidia are late and AMD do a Titan and decide WTF, the opposition have nothing to compete with so if it's good enough for Nvidia it's good enough for us. Here's your new high-end boys and girls... £1000 please. Would love to see how those who defended Titan prices respond in that scenario.

 

[Overlag]

...comprises of two GPUs. A low-end, "console-class" GPU, and a mid-range that will match/slightly beat a Fury.

you do realise that the Nano, Fury and Fury X are all just "one GPU"?

One GPU could mean one with 8192 GCN cores, and 3-4more with mutlples of 1024 disabled for instance...

(numbers made up but the point still stands).

 

[FoxEye]

It doesn't say anything in that at all about 2017 releases or what size they are or other GPU's yet to tape out.

Just that there will be 2 GPU's in 2016, 2 GPU's could be two different architectures. like we currently have in GCN 1.1 and GCN 1.2 and they spread across Low to Enthusiast end.
One with and one without HBM?

Firstly, they said "two new gpus" for 2016. That does not mean that we won't see some rebrands to fill out the range.

These new GPUs are the latest iteration of GCN (1.4 or v4). They will not be different generations/ iterations (ie one 1.4 and one 1.5). AMD have never done this before and there is no reason at all to believe this is credible. The current range only has GCN 1.0 to 1.3 parts because of rebrands.

There is also a limit to the amount of discreet cards you can make with one GPU (by fusing off parts or salvaging). It is a matter of economics. You can't make a really powerful (and expensive) chip and then make all of your mid-range and budget cards by disabling bits of it. Not cost effective.

All we know is that one "new GPU" will be console class (low-end).

One will be the "high end"... but it's unlikely to be as big as Fury (as dm has already explained). The full/best performance of this one is likely to match or somewhat exceed current Fury. "Successor to Fury" - that is wccf's words I believe, not AMD's.

But there is absolutely no evidence that a Fury +50% card is coming this year. None.

AMD have talked up the power efficiency of the small Polaris, not it's performance. Expecting this small chip to make a good mid-range card is bonkers. They called it a "power sipping" GPU for notebooks. AMD's words. It's going to be "console class". A 290/390 it aint.

The medium Polaris can't fill the whole range between Fury+ and small Polaris, you wouldn't have thought. Not if we're going to have the same amount of cards in the range as we've had up till now.

But then there's also the possibility that the range will be much smaller. Two or three cards from each GPU.

And of course you can never 100% discount the possibility of some rebrands. A 290/390 could fill the mammoth gap between the "power sipping notebook GPU" and the "high end" GPU.

 

[Jamielicious]

I always thought it was common to launch with 2 GPUs. Then you tweak with memory bus, capacity, GPU and memory speed and that gives 7-8 products.

I expect the same here. I also suspect the best price/performance cards will again be in the £280 range or the £140 range with little in between.

 

[FoxEye]

I always thought it was common to launch with 2 GPUs. Then you tweak with memory bus, capacity, GPU and memory speed and that gives 7-8 products.

For the last two generations AMD have launched only 1 new GPU. Fiji made 3 cards and Hawaii/Grenada made 2.

GP204 made 2 cards (970/980). GM100 made 2 cards (Titan/980ti).

When has 1 GPU made 4 cards? Because that's what 2 GPUs making 7-8 cards would mean.

 

[CAT-THE-FIFTH]

For the last two generations AMD have launched only 1 new GPU. Fiji made 3 cards and Hawaii/Grenada made 2.

GP204 made 2 cards (970/980). GM100 made 2 cards (Titan/980ti).

When has 1 GPU made 4 cards? Because that's what 2 GPUs making 7-8 cards would mean.

Nvidia's entire range from around £120 to £450 is served by two GPUs - the GM206 and GM204.

 

[FoxEye]

Nvidia's entire range from around £120 to £450 is served by two GPUs - the GM206 and GM204.

https://en.wikipedia.org/wiki/GeForce_900_series#GeForce_900_.289xx.29_series

GM206 made two cards and GM204 made two cards (and GM200 made two cards).

So you have 3 GPUs and 6 cards in the range. OK.

With 2 new AMD GPUs coming in 2016, that will make 4-5 cards. The thing is, the best Polaris (full fat) is going to beat FuryX but not blow it out of the water, and the worst Polaris (full fat) is "console class" - aka 1080p at 30FPS.

So with the little one all you can do is make successively worse cards than "console class" by cutting it down.

That leaves a lot of room for one GPU to fill.

/shrug. Maybe they can stick two of the little Polaris chips onto one card...

 

[CAT-THE-FIFTH]

https://en.wikipedia.org/wiki/GeForce_900_series#GeForce_900_.289xx.29_series

GM206 made two cards and GM204 made two cards (and GM200 made two cards).

So you have 3 GPUs and 6 cards in the range. OK.

With 2 new AMD GPUs coming in 2016, that will make 4-5 cards. The thing is, the best Polaris (full fat) is going to beat FuryX but not blow it out of the water, and the worst Polaris (full fat) is "console class" - aka 1080p at 30FPS.

So with the little one all you can do is make successively worse cards than "console class" by cutting it down.

That leaves a lot of room for one GPU to fill.

/shrug. Maybe they can stick two of the little Polaris chips onto one card...

Emm,its simple:
1.)GM206 filled the £100 to £200 range
2.)GM204 fill the £250 to £450 range.

So ultimately,I don't see what the problem is!

You would have smaller Polaris filling the £100 to £200 range and larger Polaris filling the £250 to £450 range.

Remember,Anandtech only estimated the size of around 120MM2 by looking at the finger of the AMD guy. It could easily be closer to 150MM2.

Plus remember,Nvidia themselves class the GTX960 as kind of console level/MOBA market and its easily faster than any of the console GPUs.

Edit!!

MOBA means games like LoL and DOTA2.

FFS,those games will play on an Intel IGP.

So you have to beware of what "markets" AMD and Nvidia say their GPUs are for.

 

[Mtom]

Thats what i'm expecting (from both sides)
Replacements in the GM206/204 - Trinidad/Tonga/Hawaii price segment first (all of them with more grunt, around +20-30%) then bring the big chips later.

Which is sad because i don't want to change my 390 for a few % more.

 

[CAT-THE-FIFTH]

It makes me wonder whether AMD is trying to go back to the old ATI ways of the "small die strategy" which means they would be more effective at fighting a price war with Nvidia. Nvidia pretty much used the same strategy during Maxwell and ATI used it for the HD3000,HD4000 and HD5000 series to fight Nvidia reasonably sucessfully.

 

[Mtom]

https://en.wikipedia.org/wiki/GeForce_900_series#GeForce_900_.289xx.29_series

GM206 made two cards and GM204 made two cards (and GM200 made two cards).

So you have 3 GPUs and 6 cards in the range. OK.

With 2 new AMD GPUs coming in 2016, that will make 4-5 cards. The thing is, the best Polaris (full fat) is going to beat FuryX but not blow it out of the water, and the worst Polaris (full fat) is "console class" - aka 1080p at 30FPS.

So with the little one all you can do is make successively worse cards than "console class" by cutting it down.

That leaves a lot of room for one GPU to fill.

/shrug. Maybe they can stick two of the little Polaris chips onto one card...

Just pointing out that we don't know if it was a full small chip or already the smallest cut down piece.

BTW the gap seems to big between the console performance and the enthusiast type GPU, there should be one more between, or a more cut down version of the big chip.