AMD 14nm FinFET Polaris GPU Size Leaks Out – 232mm² Large Die

[looncraz]

10 hours ago, Agost said:

Well, considering the jump we had with the transition from 40nm to 28nm, I wouldn't expect a 232mm^2 Polaris chip to be much more powerful than a 596mm^2 Fiji one.

I'm assuming that the 232mm^2 one is Polaris 11, not 10, since the image shows a very big difference between the two - and P11 can't be ~800mm^2

 

What image?

 

[rattacko123]

3 hours ago, looncraz said:

 

What image?

 

this one:

[Coaxialgamer]

7 hours ago, Deletive said:

Meh they can call it whatever they want and say whatever they want but if there isn't a big GCN architecture change i still won't be interested. 

Yeah,  i was also pretty frustrated when they called it gcn 2.0.

I really hope it is different from current gcn.  But from what they said,  it shoulf be a complete overhaul.  Now let's just hope they deliver. 

[looncraz]

6 hours ago, rattacko123 said:

this one:

 

I see, that's one heck of a size difference (2:1).  So if Polaris 10 is 15x15 (~225mm2), that would make Polaris 11 over 900mm2.  Which is unrealistic, to say the least.  So, if 232mm2 is accurate for Polaris 11 (15.23mm/side), then Polaris 10 is 7.62mm/side (I *love* that number :p).. which is only 58mm2.

 

If we convert to 14nm LPP density (about 2.5x higher than 28nm HPP), we get the following die-size equivalents:

Polaris 10: 145mm2

Polaris 11: 580mm2

 

These numbers are interesting in how well they line up with current products.  The Fiji die is 596mm2.  With a new version of GCN, not just a shrink, we should expect higher performance by area.  Maybe 15%... maybe even 30%.  If we extrapolate that to known products, that would be a 15~30% higher performance over Fiji for Polaris 11, but using less than half the power.  If we assume that AMD made some efforts to free up the clockrate, additional performance would come from there.  And we know that performance scales quite well with HBM, so HBM2 should bring another 5%.

 

These little things push us up to a ~50% improvement over Fiji, while still using less power than a GTX 970.  That's a pretty nice increase... and AMD can always make a bigger die when the yields improve.

 

For Polaris 10, using the same method, we see a GPU that does not match the R9 380, but falls about a third short.  Which is right around the GTX950 area of performance... which, sadly, makes sense given AMD's decision to compare it to the GTX950. This leaves the entire middle range of performance out cold.  Polaris 10 will not scale up to reach R9 380 levels, nor will Polaris 11 scale down to 390 levels (without a LOT of die harvesting).

 

Of course, this is all linear math, which is far from accurate when considering more relative area of the smaller GPU will be taken up by supporting circuitry than the larger GPU.. but that only goes to hurt the little GPU more... or, as we know Polaris 10 can keep pace with GTX 950, may well help to push Polaris 11 even higher in the performance category.

 

Of course, this is based off the graphic you showed, which may well not be even close to accurate, and a 'leaked' die size for an unknown 14nm GPU we assume to be Polaris 11... so it's not worth the pixels with which it is written.

[looncraz]

4 hours ago, Coaxialgamer said:

Yeah,  i was also pretty frustrated when they called it gcn 2.0.

I really hope it is different from current gcn.  But from what they said,  it shoulf be a complete overhaul.  Now let's just hope they deliver. 

GCN Has been extremely good to AMD.  It has kept them competitive for many years and has only just scaled to near its limits.  Following the same principal design - with refinements - makes sense, especially as that architecture's full capabilities are just now beginning to be exploited by DirectX 12.

 

They will probably improve per-CU performance, clockrate headroom, the scalar units (probably doubling their performance will be needed to head into the future), and a next generation of ACE and asynchronous shaders - just to keep nVidia in the mirror.  That would be an outsized level of tessellation improvement, doubling-down where GCN excels, and addressing its weak points.  They'd also have to ensure that it can scale beyond 64CU, or make each CU twice as fast... we've hit the point of diminishing returns for each new CU already.

[Watermelon Guy]

1 hour ago, looncraz said:

They'd also have to ensure that it can scale beyond 64CU, or make each CU twice as fast... we've hit the point of diminishing returns for each new CU already.

Yeah, they really need to make each CU better, rather than putting more CU's in. 

[Coaxialgamer]

1 hour ago, looncraz said:

GCN Has been extremely good to AMD.  It has kept them competitive for many years and has only just scaled to near its limits.  Following the same principal design - with refinements - makes sense, especially as that architecture's full capabilities are just now beginning to be exploited by DirectX 12.

 

They will probably improve per-CU performance, clockrate headroom, the scalar units (probably doubling their performance will be needed to head into the future), and a next generation of ACE and asynchronous shaders - just to keep nVidia in the mirror.  That would be an outsized level of tessellation improvement, doubling-down where GCN excels, and addressing its weak points.  They'd also have to ensure that it can scale beyond 64CU, or make each CU twice as fast... we've hit the point of diminishing returns for each new CU already.

Yeah,  i can't argue that gcn was a very durable and scalable architecture.  I just hated to think of them rehashing an architecture from 2011 on a new process node.  But it seems they will completely change it. I would really like them to improve tesselation and per CU performance,  which they probably will. 

[patrickjp93]

2 hours ago, looncraz said:

They'd also have to ensure that it can scale beyond 64CU, or make each CU twice as fast... we've hit the point of diminishing returns for each new CU already.

 

34 minutes ago, Watermelon Guy said:

Yeah, they really need to make each CU better, rather than putting more CU's in. 

 

30 minutes ago, Coaxialgamer said:

Yeah,  i can't argue that gcn was a very durable and scalable architecture.  I just hated to think of them rehashing an architecture from 2011 on a new process node.  But it seems they will completely change it. I would really like them to improve tesselation and per CU performance,  which they probably will. 

Ah the never-ending debate of moar coars vs. better coars...

 

Tessellation will have to improve drastically. It's a necessity for many high-end effects. As per clock speed, not when we're pursuing efficiency at the same time. 1100 MHz seems to be a sweet spot.

[DocSwag]

5 hours ago, patrickjp93 said:

 

 

Ah the never-ending debate of moar coars vs. better coars...

 

Tessellation will have to improve drastically. It's a necessity for many high-end effects. As per clock speed, not when we're pursuing efficiency at the same time. 1100 MHz seems to be a sweet spot.

Definitely agree they need to improve tessellation.

GCN 1.2 was actually pretty good at it. If you look at Tessmark the r9 380 and 380x both beat the gtx 960 at it. They even beat the r9 390.

They were right where they should be with it-they were better than the gtx 960 but worse than the 970.

Unfortunately though the 980 ti still beats the fury x.

They did say they redesigned the geometry processor so that should be interesting.

[AlTech]

On 2/14/2016 at 1:44 AM, Watermelon Guy said:

There saying this is equal in density to 450mm in 28nm

That's slightly larger than a 390X....................Damn. That's huge. I hope this has 12.4 Billion transistors then (double R9 390X's transistor count).

 

I hope Polaris 11 will be awesome. If it is then i might upgrade. and Get a Fury X successor or Fury Successor depending on pricing. Hopefully Polaris 11 will be around $500 and not $700. 

[AlTech]

13 hours ago, patrickjp93 said:

 

 

Ah the never-ending debate of moar coars vs. better coars...

 

Tessellation will have to improve drastically. It's a necessity for many high-end effects. As per clock speed, not when we're pursuing efficiency at the same time. 1100 MHz seems to be a sweet spot.

Sometimes it's better for power efficiency to have more cores which are clocked slower than less core clocked higher.

 

for example. If i were AMD i might make a 900MHz GPU with 4352 stream processors (for example) instead of a 1100MHz GPU with 4096 stream processors.

 

The former may consume less power depending on what architecture was used.

[Watermelon Guy]

3 hours ago, AluminiumTech said:

Sometimes it's better for power efficiency to have more cores which are clocked slower than less core clocked higher.

 

for example. If i were AMD i might make a 900MHz GPU with 4352 stream processors (for example) instead of a 1100MHz GPU with 4096 stream processors.

 

The former may consume less power depending on what architecture was used.

Who knows, we dont know what sort of improvements gcn 1.4 (4.0) will bring. 

[AlTech]

Just now, Watermelon Guy said:

Who knows, we dont know what sort of improvements gcn 1.4 (4.0) will bring. 

More power efficiency and/or more performance probably.

[Watermelon Guy]

Just now, AluminiumTech said:

More power efficiency and/or more performance probably.

Yeah, I really hope its not just power management tricks, like fiji, to get power use down. 

[Citadelen]

23 minutes ago, Watermelon Guy said:

Yeah, I really hope its not just power management tricks, like fiji, to get power use down. 

Fiji? Dont you mean Maxwell?

[DocSwag]

1 hour ago, Watermelon Guy said:

Who knows, we dont know what sort of improvements gcn 1.4 (4.0) will bring. 

I think they will be calling it gcn 2.0, but that's only my guess.

They've actually already told us everything they redesigned: http://www.zdnet.com/article/a-closer-look-at-amds-polaris-graphics/

we just don't know exactly what they did to each component.

Personally I find the most interesting part the geometry processor. AMD has always been behind in tesselation, and this could really help. They already had some big improvements with gcn 1.2 (the r9 380 actually beat the GTX 960 at it) but the fury X still trailed the 980 ti. Also they are redesigning the CUs which could be interesting.

[Watermelon Guy]

1 hour ago, Citadelen said:

Fiji? Dont you mean Maxwell?

Maxwell is nvidia. Fiji is the amd flagship, which will be replaced this year by polaris or mayber vega. 

32 minutes ago, DocSwag said:

I think they will be calling it gcn 2.0, but that's only my guess.

They've actually already told us everything they redesigned: http://www.zdnet.com/article/a-closer-look-at-amds-polaris-graphics/

we just don't know exactly what they did to each component.

Personally I find the most interesting part the geometry processor. AMD has always been behind in tesselation, and this could really help. They already had some big improvements with gcn 1.2 (the r9 380 actually beat the GTX 960 at it) but the fury X still trailed the 980 ti. Also they are redesigning the CUs which could be interesting.

i think there officially calling it gcn 4.0. 

7970 1.0

290x 2.0

Fury x 3.0

Polaris 4.0

This press just called 2, and 3 1.1 and 1.2 because there isnt much difference.  The fury x trailed behind because it lacked enough ROP's which became a bottleneck. 

[NateGSR117]

I can't wait to see these cards in action. 

[Citadelen]

1 hour ago, Watermelon Guy said:

Maxwell is nvidia. Fiji is the amd flagship, which will be replaced this year by polaris or mayber vega. 

i think there officially calling it gcn 4.0. 

7970 1.0

290x 2.0

Fury x 3.0

Polaris 4.0

This press just called 2, and 3 1.1 and 1.2 because there isnt much difference.  The fury x trailed behind because it lacked enough ROP's which became a bottleneck. 

You said you wanted real efficiency gains, not power management tricks like Fiji, but it was Nvidia that cut out compute to give the image of efficiency. I haven't heard of Fiji doing anything like that.

[DocSwag]

1 hour ago, Watermelon Guy said:

Maxwell is nvidia. Fiji is the amd flagship, which will be replaced this year by polaris or mayber vega. 

i think there officially calling it gcn 4.0. 

7970 1.0

290x 2.0

Fury x 3.0

Polaris 4.0

This press just called 2, and 3 1.1 and 1.2 because there isnt much difference.  The fury x trailed behind because it lacked enough ROP's which became a bottleneck. 

In reality though they called the 7970 gcn 1.0, 290x gcn 1.1, and fury x 1.2.

It makes sense to call it 4th gen because it is the fourth iteration of GCN.

Looking at the above it would make sense to be 1.3, but this is the biggest change they have done to their GPUs since GCN itself so it makes sense to be 2.0.

They call it 4th generation GCN, not GCN 4.0. It still isn't certain exactly what it will be called though, so you could be right or I could. Only time will tell

[Mihle]

I just hope their top end card will be the same heat and power consuption as fury x, and use the 14nm on getting a lot better peformance.

 

I dont care about consumption as long as a 1000w psu can siport one gpu and a cpu. heat can be fixed with water cooling. what I want is peformance 

[RexinOridle]

No offence but I don't know what to expect from a company when their majority of R&D is based in China.

[Sauron]

Unfortunately this doesn't really tell us anything on its performance

[patrickjp93]

38 minutes ago, Mihle Gaming said:

I just hope their top end card will be the same heat and power consuption as fury x, and use the 14nm on getting a lot better peformance.

 

I dont care about consumption as long as a 1000w psu can siport one gpu and a cpu. heat can be fixed with water cooling. what I want is peformance 

Your wants are low down on the list compared to the needs of data centers. Better get used to that idea.

[Mihle]

10 hours ago, patrickjp93 said:

Your wants are low down on the list compared to the needs of data centers. Better get used to that idea.

is it that hard to just take their top end card where they do care about power consumption and put in more transistors to get it to perform better? (without a complete re design)