wccftechAMD Pirate Islands : R9 300 Series Alleged Specifications Detailed

csbin

Senior member
Feb 4, 2013
838
351
136
http://wccftech.com/amd-pirate-islands-r9-300-series-bermuda-fiji-treasure-islands-xtx/

*Nothing mentioned in this leak is set in concrete, AMD is bound to change a few things here and there.
[Exclusive] [Editorial]: I have just recently gotten my hands on a somewhat outdated and alleged AMD GPU specification details, the essentials of which should nonetheless hold true. The specifics mentioned in this post are liable to change and the primary focus of this leak is to serve as the basis of all future Pirate Islands speculation. This is going to be a very very interesting post, so buckle up, but keep that pinch of salt handy.


So lets begin, basically the AMD Details show three GPUs, the R9 390X, R9 380X and R9 370X, namely the Pirate Islands series. According to these details all GPUs will be based on the TSMC 20nm Node and will be true Pirate Island cores, which is why nomenclature of the dies will be derivative in nature. Another thing mentioned is that all Pirate Island GPUs will have the DirectX 12 Hardware Feature Set. According to these details the R9 370X is slated for announcement/arrival sometime this July-August this year. The R9 370X will feature a ‘Treasure Island XTX’ core and supposedly has 1536 Stream-processors, 96 Texture Units and 48 ROPs on the uncut die. Of course this road-map precedes the report we got of TSMC having a little trouble with 20nm so I am not sure how valid the time frame is anymore.
Now according to the details, we will see the R9 390X before the R9 380X. The R9 390X will feature the ‘Bermuda XTX’ core and feature a staggeringly high number of SPs, ie. 4224. The Texture Unit count is 264 with 96 ROPs. It is slated to show up late 2014 around November or so. Once again keep in mind the TSMC comment above. Finally the R9 380X which will feature the ‘Fiji XTX’ core and the following die configuration: 3072 SPs, 192 Texture Units and 72 ROPs and should appear sometime early 2015. Now interestingly the R9 390X ‘Bermuda XTX’ GPU is slated to have 512 Bits of Memory Bandwidth while as the R9 380X is slated to have 384 Bits with the R9 370X coming last with 256 Bits. Now the road-map even hints at the clock speeds of AMD’s future GPUs but I think it is prudent to mention that all these tech specs at this point in time are about as solid as the wind. However let us continue, with the help of this table:

TrG.jpg
 
Last edited:

HurleyBird

Platinum Member
Apr 22, 2003
2,678
1,266
136
I hope the specs are true, source has a shaky track record though.

Specs are technically believable, but definitely on the very edge of what can reasonably be expected. Whether or not these are accurate, a ROP increase to better cope with 4K is a sound prediction.
 
Last edited:

Gloomy

Golden Member
Oct 12, 2010
1,469
21
81
sorry this just seems insane to me, the idea that either GPU manufacturer can deliver something like that, that would be an obviously amazing jump in just one generation, and I'm just. so used to disappointment

maybe I'm a cynic

those ROPs... how do you even keep that fed? :|
 
Feb 19, 2009
10,457
10
76
Gloomy you are living up to your username. :D

Node jumps have always (as far as I recall) delivered amazing leaps in performance. Its to be EXPECTED. Anything less and I would consider it a failure.
 

TreVader

Platinum Member
Oct 28, 2013
2,057
2
0
These specs look believable, but it looks startlingly similar to the current GTX-700 series lineup. Exact same number of SPs as the 680 with an extra 12 ROPs. I'd be surprised if they stuck with a 256bit bus on the 370X considering the R7 265 has the exact same memory bandwidth (this list the 370X as having 10% LESS bandwidth than the R7 265!)
 

SlowSpyder

Lifer
Jan 12, 2005
17,305
1,001
126
If true it would seem AMD suddenly fell in love with ROP's! They stuck with 32 ROP's on their flagships for a number of generations. 1080P is becoming the new 1024x768, it would seem. I assume they're going for 4k performance.
 
Feb 19, 2009
10,457
10
76
They and NV would HAVE to go for 4K performance, these are flagship cards most likely out in 2015. We're getting 4K monitors now that cost less than the top single GPUs. In another year, 4K monitors will be as 1440p monitors are now, just only a little bit above average.

I know I am definitely gonna get a 4K monitor asap.
 

TreVader

Platinum Member
Oct 28, 2013
2,057
2
0
Actually after looking at it twice, I think it's legit. The 7870 was 1280/76/32 and this looks like just about 150% the performance of a 7870, although it might get a bit choked for bandwidth. (370X)


96 ROPs is def the min I would attach to a card meant for 4k. I was hoping for 128.
 

Bateluer

Lifer
Jun 23, 2001
27,730
8
0
I hope the specs are true, source has a shaky track record though.

Specs are technically believable, but definitely on the very edge of what can reasonably be expected. Whether or not these are accurate, a ROP increase to better cope with 4K is a sound prediction.

Huh? Seems like a fairly natural evolution of the current Hawaii design, which has 2816 SPs. 1536 and 3072 seem pretty tame for a full node jump. 4224, maybe a bit high, but it could also be another 250W+ card with hybrid cooling.
 
Feb 19, 2009
10,457
10
76
Huh? Seems like a fairly natural evolution of the current Hawaii design, which has 2816 SPs. 1536 and 3072 seem pretty tame for a full node jump. 4224, maybe a bit high, but it could also be another 250W+ card with hybrid cooling.

High would be 2816 x 2, 4224 is a very mild (50%) and tame leap. If true, it would mean VI will be a small-die approach.
 

Fire&Blood

Platinum Member
Jan 13, 2009
2,331
16
81
I'm hoping it will be too soon for 4K, for my wallet's sake. However, if 2 of these upcoming cards (no way a single GPU) can handle 4K gaming with minimal IQ compromises, I'm in.
 

Techhog

Platinum Member
Sep 11, 2013
2,834
2
26
That R9 370X looks extremely disappointing, but the the 380X would be amazing for $300. Assuming my 7950 isn't actualoly damaged and can last until then, that's what I'd upgrade to.
 

DiogoDX

Senior member
Oct 11, 2012
746
277
136
High would be 2816 x 2, 4224 is a very mild (50%) and tame leap. If true, it would mean VI will be a small-die approach.
With 20nm density 4200sps will be almost the size of the 7970(350mm2). I think is realistic for the first bacth of the 20nm.

But I remember that AMD is working whit Hynix in HBM (stacked memory).:hmm:

http://electroiq.com/blog/2013/12/amd-and-hynix-announce-joint-development-of-hbm-memory-stacks/
http://sites.amd.com/us/Documents/TFE2011_006HYN.pdf
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,686
1,221
136
Shader Engine:
1 Geometry Engine per SE
1 Rasterizer per SE
1 to 16 (GCN Cores) Compute Units per SE
1 to 4 RBEs per SE

There can up to 4 for a GCN GPU.

R9 290X / 290:
4 Shader Engines;
4 Geometry Engine / 1 per SE
4 Rasterizers / 1 per SE
40 to 44 GCN Cores / 10 to 11 per SE
16 RBEs / 4 per SE

R9 280X / 280:
2 Shader Engines;
2 Geometry Engines / 1 per SE
2 Rasterizers / 1 per SE
28 to 32 GCN Cores / 14 to 16 per SE
8 RBEs / 4 per SE

R9 270X / 270 / 265:
2 Shader Engines;
2 Geometry Engines / 1 per SE
2 Rasterizers / 1 per SE
16 to 20 GCN Cores / 8 to 10 per SE
8 RBEs / 4 per SE

R7 260X / 260:
2 Shader Engines;
2 Geometry Engines / 1 per SE
2 Rasterizers / 1 per SE
12 to 14 GCN Cores / 6 to 7 per SE
4 RBEs / 2 RBEs per SE

--- What happens when you double somethings in mirror format ---

R7 36_(x):
2 Shader Engines;
2 Geometry Engines / 1 per SE
2 Rasterizers / 1 per SE
24 to 28 GCN Cores / 12 to 14 per SE
8 RBEs / 4 RBEs per SE

1792-1536 ALUs / 112-96 TMUs / 32 ROPs

/////

R9 37_(x):
4 Shader Engines;
4 Geometry Engines / 1 per SE
4 Rasterizers / 1 per SE
32 to 40 GCN Cores / 8 to 10 per SE
16 RBEs / 4 per SE

2560-2048 ALUs / 160-128 TMUs / 64 ROPs

////

R9 38_(x):
4 Shader Engines;
4 Geometry Engines / 1 per SE
4 Rasterizers / 1 per SE
56 to 64 GCN Cores / 14 to 16 per SE
16 RBEs / 4 per SE

4096-3584 ALUs / 256-224 TMUs / 64 ROPs

These all can fit into the GCN Shader Engine ruleset which means after this point it must be the GCN2 Shader Engine ruleset.
 

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
Huh? If its designed for 20nm that's nothing super spectacular. It is a full node jump afterall.

Gloomy you are living up to your username. :D

Node jumps have always (as far as I recall) delivered amazing leaps in performance. Its to be EXPECTED. Anything less and I would consider it a failure.

High would be 2816 x 2, 4224 is a very mild (50%) and tame leap. If true, it would mean VI will be a small-die approach.

Keep in mind that yields will not be as high as 28nm, so a big die would be worse for yields. Because you pay per wafer, not per die, 2 times as much cores would be more expensive. Also, 20nm might not give a lot of power and performance increases because it doesn't have FinFETs to reduce leakage.
 

TreVader

Platinum Member
Oct 28, 2013
2,057
2
0
Keep in mind that yields will not be as high as 28nm, so a big die would be worse for yields. Because you pay per wafer, not per die, 2 times as much cores would be more expensive. Also, 20nm might not give a lot of power and performance increases because it doesn't have FinFETs to reduce leakage.

You should keep in mind that the prices for GPUs have risen substantially, and it would pay for AMD to release a big-die (500mm+) card like the 390X if nvidia is going small die with GM104.


If the specs are right, nvidia will lose the single gpu crown, after already losing mGPU and dual-GPU cards. The value of AMD being considered the fastest is substantially more than whatever margin they might lose with a big die. It looks like AMD might pull off another 9700pro moment.

That said, I doubt ALL the numbers are accurate. They probably aren't just made up out of thin air.