AMD Pirate Islands Radeon R9 300 Series

[humbug]

Grain of salt as its wccftech.

Source
http://wccftech.com/amd-pirate-islands-r9-300-series-bermuda-fiji-treasure-islands-xtx/?

R9 390X 4224 SP's, thats 50% more than the R9 290X, and 96 ROP's?????? wow

 

[James J]

R9 390X 4224 SP's, thats 50% more than the R9 290X, wow

I think the new cards need to have such a performance increase over the current gen stuff tbh or people will seriously think twice about upgrading. Big performance gap between a 7970 and a 290x, same with a 680 and 780 card from the other camp. I wouldn't upgrade my gpus unless the speed improvement was there and useful to me in the games I play.

I also hope that the GPU mining craze has died off by then and AMD cards have come back down to their usual good price points.

 

[humbug]

I think the new cards need to have such a performance increase over the current gen stuff tbh or people will seriously think twice about upgrading. Big performance gap between a 7970 and a 290x, same with a 680 and 780 card from the other camp. I wouldn't upgrade my gpus unless the speed improvement was there and useful to me in the games I play.

I also hope that the GPU mining craze has died off by then and AMD cards have come back down to their usual good price points.

I just bought a 290, i dread it being out of date and slow in less than a year

 

[Boomstick777]

I just bought a 290, i dread it being out of date and slow in less than a year

It won't be out of date or slow. Nvidia and AMD know how to market things to keep us suckers always wanting the latest Shiny new thing

Determined to not be an early adopter this time around. More interested in the second wave of 20nm cards. Hopefully AMD or Nvidia will have a single GPU by then that can match GTX 780 SLI and be a proper upgrade.

 

[Freddie1980]

I wonder if AMD will get thermals back under control with 20nm? I know thermals aren't just about power draw but its a big factor. In years gone by their cards have been less power hungry then the green teams equivalent product but since 28nm Nvidia have stolen the performance per watt crown from AMD (at least at sensible resolutions 1080, 1440 etc). AMD should also improve thier reference boards I noticed that the Sapphire 290x trix used at least 30 watts less under load then a stck 290x (see Anandtech).

 

[wazza300]

 

[Kaapstad]

I just bought a 290, i dread it being out of date and slow in less than a year

It will still be a fast card in a years time and don't be surprised if it gets rebadged as a 370X, another reason why I think the table in the OP is dodgy.

 

[Kaapstad]

It won't be out of date or slow. Nvidia and AMD know how to market things to keep us suckers always wanting the latest Shiny new thing

Determined to not be an early adopter this time around. More interested in the second wave of 20nm cards. Hopefully AMD or Nvidia will have a single GPU by then that can match GTX 780 SLI and be a proper upgrade.

+1

The second wave will be the cards to have, if you apply that to 28nm it is the difference between a GTX680 and a 780ti.

 

[MrMarvelous]

Not too bothered. No games are really pushing my 290, I think i'll be keeping it for a while.

 

[lltfdaniel]

That is what i am thinking which is about games not pushing the limits of the gpu.

I have the 280x which is same as 7970 ghz.

 

[jakspyder]

+1

The second wave will be the cards to have, if you apply that to 28nm it is the difference between a GTX680 and a 780ti.

That second wave is hopefully what is going to push 4k resolution into the forefront.

 

[triss]

It won't be out of date or slow. Nvidia and AMD know how to market things to keep us suckers always wanting the latest Shiny new thing

Determined to not be an early adopter this time around. More interested in the second wave of 20nm cards. Hopefully AMD or Nvidia will have a single GPU by then that can match GTX 780 SLI and be a proper upgrade.

Really Boomie you know when its released the itch will get to u

 

[almighty15]

72 and 96 ROPS.... That's what we need if 4k is to become playable on a single card.

 

[Freddie1980]

+1

The second wave will be the cards to have, if you apply that to 28nm it is the difference between a GTX680 and a 780ti.

No guarantees a second wave will bring the big leaps your hoping for. Remember 40nm? 5780/6790 & 480/580 the improvements where about 10% between product lines.

 

[almighty15]

No guarantees a second wave will bring the big leaps your hoping for. Remember 40nm? 5780/6790 & 480/580 the improvements where about 10% between product lines.

In tessellated games the difference between a 5870 and 6970 was way more then 10%

 

[kitch9]

None of these or the new maxwells have taped out yet so July/August would be pushing it I reckon.

It depends what they find when they tape out as well. One thing I know, the smaller the process, the more the laws of physics pushes back.

 

[drunkenmaster]

Total rubbish in this table

If you look at the numbers there is no architecture scaling and why do the clockspeeds increase as the chips get bigger lol.

Neither of these things are indicators.

First with clock speeds, voltage dictates stable clock speed, voltage dictates power, power budget on cheaper cards are important, high end card, it uses what power it uses. HIgh end card buyers aren't going to refuse a card because it uses 280 instead or 250W, but Dell will happily reject buying 5million lower end cards because they use 100W instead of 90W.

Clock speed is determined by power budget, not what the chips can actually do but the target market they are aiming at.

In terms of architecture, who knows where it scales. The bigger the chip the more likely to have disabled units for yields, likewise AMD are really attacking the professional/compute market more heavily and seeing significant increase in overall margins because of it. Comparatively Nvidia managed to make a fairly competitive smaller chip than AMD by removing most real compute ability. 99% of gpu buyers don't care about DP performance in the slightest. Meaning at some stage I wouldn't be surprised if AMD cut some compute stuff out of everything but the high end gpu to save power and die size as Nvidia did. So there are two pretty large reasons for a high end architecture to not scale precisely with lower end ones.


not saying these are real or accurate leaks at all, just that clock speeds and tmu's/rops don't need to scale or be in a specific ratio throughout a product range.


If they were accurate(not saying they are) then it would hint at rops being in groups of 12.

Based on Hawaii I would say there were 12 CU's per shader engine, 64 shaders per cu and, 4tmu's per CU(as currently) but with one extra CU per engine(currently 11... strange number to have frankly).

So a midrange card would simply have 2 or 4 shader engines, the high end 6 shader engines with maybe one cu disabled per engine for yields giving 4608 shaders total with 6x64 disabled for 4224 shaders usable. I wouldn't be surprised if with maybe most new generations the way compute is being used in general and in gaming, that we don't get further splitting of the gpu into compute units. So maybe it will be 8 shader engines in the midrange and 12 in the high end. The more it's split the more different things can be worked on in a particular cluster at the same time.

It would give it effectively very very similar layout to what Hawaii is currently in terms of tmu's/shaders per CU, cu per engine rops vs number of engines. Rops are "scalable" as they aren't included in any specific part of the core but on the back end so they can effectively chose how many they want for a given product. With Hawaii when discussing the architecture they said effective "up to X number of rops" because they can really chose by the product how many rops to stick in.

 

[Doco]

With one gpu recently gone faulty on me - looks like I could have something to lookfoward to at the end of the year (so start saving now lol)

I'l keep an eye out for any further leaks - on either side

 

[humbug]

Neither of these things are indicators.

First with clock speeds, voltage dictates stable clock speed, voltage dictates power, power budget on cheaper cards are important, high end card, it uses what power it uses. HIgh end card buyers aren't going to refuse a card because it uses 280 instead or 250W, but Dell will happily reject buying 5million lower end cards because they use 100W instead of 90W.

Clock speed is determined by power budget, not what the chips can actually do but the target market they are aiming at.

In terms of architecture, who knows where it scales. The bigger the chip the more likely to have disabled units for yields, likewise AMD are really attacking the professional/compute market more heavily and seeing significant increase in overall margins because of it. Comparatively Nvidia managed to make a fairly competitive smaller chip than AMD by removing most real compute ability. 99% of gpu buyers don't care about DP performance in the slightest. Meaning at some stage I wouldn't be surprised if AMD cut some compute stuff out of everything but the high end gpu to save power and die size as Nvidia did. So there are two pretty large reasons for a high end architecture to not scale precisely with lower end ones.


not saying these are real or accurate leaks at all, just that clock speeds and tmu's/rops don't need to scale or be in a specific ratio throughout a product range.


If they were accurate(not saying they are) then it would hint at rops being in groups of 12.

Based on Hawaii I would say there were 12 CU's per shader engine, 64 shaders per cu and, 4tmu's per CU(as currently) but with one extra CU per engine(currently 11... strange number to have frankly).

So a midrange card would simply have 2 or 4 shader engines, the high end 6 shader engines with maybe one cu disabled per engine for yields giving 4608 shaders total with 6x64 disabled for 4224 shaders usable. I wouldn't be surprised if with maybe most new generations the way compute is being used in general and in gaming, that we don't get further splitting of the gpu into compute units. So maybe it will be 8 shader engines in the midrange and 12 in the high end. The more it's split the more different things can be worked on in a particular cluster at the same time.

It would give it effectively very very similar layout to what Hawaii is currently in terms of tmu's/shaders per CU, cu per engine rops vs number of engines. Rops are "scalable" as they aren't included in any specific part of the core but on the back end so they can effectively chose how many they want for a given product. With Hawaii when discussing the architecture they said effective "up to X number of rops" because they can really chose by the product how many rops to stick in.

I think its about time AMD did what Nvidia do, is 5.6 TeraFlops necessary on a gaming card?

The R9 390X looks like it would have about 9 TeraFlops of DP Compute, that's insane. AMD could cut half the compute performance out, save on a lot of power consumption and still end up with huge compute performance.

 

[LtMatt]

AMD Pirate Islands Can be Announced The Summer

In October 2014 will be released Radeon R9 390X (Bermuda XTX) with 4224 shader and 512-bit memory bus. Frequency formula will be as follows: 1 GHz GPU chip and 7 GHz for memory chips. Later on, in early 2015, debuts of the Radeon R9 380X, aka Fuji XTX. In the arsenal of new items will be 3072 stream processors, 384-bit memory bus and chips GDDR5.Exemplary operating frequencies R9 380X make 900/6000 MHz.

Spoiler

Source
http://extremespec.net/amd-pirate-islands-announced-summer/