Polaris 10 and 11 confirmed to be GDDR5 based

[railven]

Raja Koduri has already stated that shading efficiency and throughput has increased significantly.

http://venturebeat.com/2016/01/15/a...-to-full-graphics-immersion-with-16k-screens/

I am quite optimistic about the full Polaris 10 SKU (2560 sp) beating Fury X.

Hell, I think Hawaii/Grenada with more ROPs would beat Fiji XT.


I'm still on the fence. I got my hopes up for Fury X and was left sorely disappointed. Hopefully AMD can turn that ship around and make their flagship desirable. All the selling points regurgitated around here did nothing for the Fiji-line up. Hawaii/Grenada stole the show.

 

[Head1985]

Another speculation

Polaris 10 should be 470 series and 11 should be 460series.
Polaris10 full=470x GTX980 performance Much less TDP-120w 2560SP 256bit DDR5 8GBram. - 7870 replacement.
Polaris10 cutdown=470 Faster than GTX970 by 5-8% much less TDP-100w 2304SP 256Bbit DDR5 4-8GB. -7850 replacement
Polaris 11 Full=460x GTX960 performance up to + 5% much less TDP-50-60w 1280SP 128bit 4GB DDR5 - 7770 replacement
Polaris11 cutdown-460 GTX950 performance Up to 5% Much less TDP-40-50w 1024SP 128Bit 4GB DDR5- 7750 replacement


Vega11 full=480x faster than GTX980TI by 20-30% much less TDP-180w 3840SP HBM2 8GB - 7970 replacement
Vega11 cutdown=480 faster than GTX 980TI by 10-12% much less TDP-160w 3328SP HBM2 8GB - 7950 replacement
Vega10 full=490x faster than GTX980TI by 60% TDP 225w 5120SP 8-16GB HMB2 -290x/390x replacement
Vega10 cutdown=490 faster than GTX980TI by 50% TDP 200w 4608SP 8-16GB HBM2- 290/390 replacement

 

[railven]

Man, if your prediction has any hints of true, it makes me sad that I'd have to wait until 2017 to upgrade

And no way in hell I'm touch mGPU's again. Tried it from both camps, and they both sucked, and it seems now it's even worse.

 

[Vesku]

Polaris 10 looking to be somewhere around 390X or Fury performance at 150W or a bit less. I suppose if it's $330 or less it will be AMD's version of the 970, solid card for those 2 or more GPU generations behind. Makes sense with Polaris 11 preview portraying a next gen 750 TI-esque product.

I'm thinking part of whether Polaris 10 has a good reception will be whether AMD adopts a more advanced default boost/turbo, and the better review game charts that would result, in addition to where it lands price/perf wise. I expect power efficiency to be a relative non-issue this generation between Nvidia and AMD. Posts about how important power consumption is for GPUs are already dropping in frequency in popular online forums. Nvidia viral marketing will have to find something else to play up.

Odds are good Nvidia will have the higher performing card for 2016 but that might also mean a poor price/perf ratio. Pascal may also show some weakness in games that fully leverage and exploit Async Compute (not something we'll see in most or any Nvidia Gameworks titles, though).

 

[Paratus]

Polaris 10 looking to be somewhere around 390X or Fury performance at 150W or a bit less. I suppose if it's $330 or less it will be AMD's version of the 970, solid card for those 2 or more GPU generations behind. Makes sense with Polaris 11 preview portraying a next gen 750 TI-esque product.

I'm thinking part of whether Polaris 10 has a good reception will be whether AMD adopts a more advanced default boost/turbo, and the better review game charts that would result, in addition to where it lands price/perf wise. I expect power efficiency to be a relative non-issue this generation between Nvidia and AMD. Posts about how important power consumption is for GPUs are already dropping in frequency in popular online forums. Nvidia viral marketing will have to find something else to play up.

Odds are good Nvidia will have the higher performing card for 2016 but that might also mean a poor price/perf ratio. Pascal may also show some weakness in games that fully leverage and exploit Async Compute (not something we'll see in most or any Nvidia Gameworks titles, though).

I'm not so sure. Since I've been playing with transistor density this morning AMD hit higher densities than NV did.

Fiji: 14.9 million per mm^2
980TI: 13.3 million per mm^2

12% more transistors for the same area. I think in the previous and current generations that has translated to more feature support (eg ACEs) but less performance when those features were unused for DX11.

Now we are moving towards DX12 and AMD may not need to spend as many transistors on features since their architecture already supports most of it. They can spend their bonus 12% on performance.

Just a thought.

 

[Vesku]

I'm not so sure. Since I've been playing with transistor density this morning AMD hit higher densities than NV did.

Fiji: 14.9 million per mm^2
980TI: 13.3 million per mm^2

12% more transistors for the same area. I think in the previous and current generations that has translated to more feature support (eg ACEs) but less performance when those features were unused for DX11.

Now we are moving towards DX12 and AMD may not need to spend as many transistors on features since their architecture already supports most of it. They can spend their bonus 12% on performance.

Just a thought.

I just think Nvidia's mid-range GDDR5 die will be a bit bigger than AMD's. They have higher expectations from customers in regards to 1080 performance vs 980 and less pressure regarding perf/watt, imo, and I'm thinking that will mean something a bit closer to 300mm2, 250-280 perhaps.

 

[96Firebird]

According to some random german guy in a forum...

This is nothing new, people on this very forum trust random forum users and their posts... D:

And I'm actually a little surprised more people aren't taking this news badly, considering the posts in this thread. They were almost speaking as if the new cards aren't using HBM, they're trash!

 

[Head1985]

some people just expecting too much.
polaris 10 is same tier as GP106/1060 or GTX960 successor.

Polaris 10(232mm2) is 7870(212mm2) successor and that GPU is slighly slower than GTX960 now.

So people really expecting GTX960 successor BEAT TITANX.Because its just what they are saying with polaris 10 and fury X.
GTX980 performance is more real.We talking about 80-90% performance increase.

GTX980 is now 82% faster than GTX960 in 1080P
390x is now 92% faster than 7870/270x in 1080P
FUry x is now 120% faster than 7870/270x in 1080P

So what is more real?7870 succesor have 390x/GTX980 performance or 7870successor will have 120% more performance?

 

[jpiniero]

Somewhere in between the 390X and Fury X. Which I expect would be competitive with the cut GP104.

 

[ShintaiDK]

This is nothing new, people on this very forum trust random forum users and their posts... D:

And I'm actually a little surprised more people aren't taking this news badly, considering the posts in this thread. They were almost speaking as if the new cards aren't using HBM, they're trash!

Reality is always a bitter pill when you hype

GDDR5(X) will rule for a long time in the classes below flagship.

 

[ShintaiDK]

Somewhere in between the 390X and Fury X. Which I expect would be competitive with the cut GP104.

Unless the cards are fit with GDDR5X. Then the 512bit bus of Hawaii will just be too much for a 256bit one in average. Just like Tahiti did with Tonga. And that was with a much smaller margin in bandwidth.

 

[itsmydamnation]

Unless the cards are fit with GDDR5X. Then the 512bit bus of Hawaii will just be too much for a 256bit one in average. Just like Tahiti did with Tonga. And that was with a much smaller margin in bandwidth.

You cant say that with any authority, increase in Dram clock, increase in memory compression, increase in cache structures, primitive discard engine could all go hand it hand to make up the deficit.

 

[AtenRa]

Polaris 10 and 11 are made for 1080p and 1440p. Polaris 10 could reach and even overcome in some games Hawaii performance at 1080/1440p, at 4K Fiji will be faster until Vega arrives in early 2017.

 

[ShintaiDK]

You cant say that with any authority, increase in Dram clock, increase in memory compression, increase in cache structures, primitive discard engine could all go hand it hand to make up the deficit.

You can keep on hoping.

DRAM clock is the same if we are to believe the leaks. So its half the bandwidth.

ROV and CR is already supported for Maxwell/Skylake. So even for that to work, you need people to code for DX12.1 feature level first.

 

[itsmydamnation]

You can keep on hoping.

DRAM clock is the same if we are to believe the leaks. So its half the bandwidth.

im not hoping anything, you are making assumptions, what are my assumptions?

Hawaii is almost never bandwidth bottlenecked (even the 290 @ 1250), it doesn't need a 512bit bus, the reason for 512bus could have just been about adding more ROP's.

Also Polaris in the leak is only clocked at 800mhz so its likely both core and memory are clocked lower in these samples. Also the sample looks like its the cutdown not full Polaris 10.

 

[Vesku]

You can keep on hoping.

DRAM clock is the same if we are to believe the leaks. So its half the bandwidth.

ROV and CR is already supported for Maxwell/Skylake. So even for that to work, you need people to code for DX12.1 feature level first.

I don't think those clocks will be retail clocks. Wasn't the core at 800MHz? One of the AMD comments during Polaris sneak peak was they designed for higher clocks than 28nm GCN.

 

[Mahigan]

im not hoping anything, you are making assumptions, what are my assumptions?

Hawaii is almost never bandwidth bottlenecked (even the 290 @ 1250), it doesn't need a 512bit bus, the reason for 512bus could have just been about adding more ROP's.

Also Polaris in the leak is only clocked at 800mhz so its likely both core and memory are clocked lower in these samples. Also the sample looks like its the cutdown not full Polaris 10.

Hawaii is memory bandwidth bottle necked.

But Polaris is not Hawaii.
- Increased L2 Cache
- Instruction pre-fetching
- Primitive Discard Acceleration
- New color compression algorithms
- New memory controller
- Most likely GDDR5x usage.

This means that Polaris won't be memory bandwidth bottlenecked. Instruction prefetching allows Polaris to pull instructions out of memory and into cache (like textures for example) thus alleviating memory bandwidth bottlenecks.

The increased L2 Cache is to help with storing these new instructions.

Primitive Discard acceleration is to help in getting rid of un-needed triangle/rasterization data (culling). Alleviating the load on cache and memory bandwidth.

And the new memory controllers + color compression + GDDR5x means that Polaris shouldn't be constrained by memory bandwidth like Hawaii was.

 

[itsmydamnation]

that graph doesn't show 290 as being bottle necked by bandwidth.

but also GCN currently the ROP cache is separate from the "not quite unified, unified L2". so making a true unified L2 (like maxwell) can help off chip requirements as well.

edit: i think polaris is unlikely to use GDDR5x its to late.

 

[Silverforce11]

232mm2 is pretty small to expect Fury X+ performance, but it's new advantages may mean the minimum FPS could exceed Fury X in CPU and GPU bound scenarios.

Otherwise, I would guess it falls around 390X levels.

NV's mid-range Kepler was much bigger, and mid-range Maxwell was massive in comparison. But I would expect GP104 to be around 300mm2.

Async or not, Polaris 10 is too small to compete with that.

 

[Mahigan]

that graph doesn't show 290 as being bottle necked by bandwidth.

but also GCN currently the ROP cache is separate from the "not quite unified, unified L2". so making a true unified L2 (like maxwell) can help off chip requirements as well.

edit: i think polaris is unlikely to use GDDR5x its to late.

Here's the full image. Read the article bellow it.

 

[AtenRa]

Why people believe GP104 will be 300mm2 ?? do we have any leak for that size ?? Because if Polaris 10 will be 230mm2, im expecting GM104 at 250mm2 range.

 

[Silverforce11]

Why people believe GP104 will be 300mm2 ?? do we have any leak for that size ?? Because if Polaris 10 will be 230mm2, im expecting GM104 at 250mm2 range.

NV doesn't operate on the same basis as AMD.

They have Teslas and Quadros to make the bulk of their profits. These markets expect big performance gains to justify upgrades of major infrastructure.

 

[jpiniero]

Why people believe GP104 will be 300mm2 ?? do we have any leak for that size ?? Because if Polaris 10 will be 230mm2, im expecting GM104 at 250mm2 range.

That's what I am thinking too.

They have Teslas and Quadros to make the bulk of their profits. These markets expect big performance gains to justify upgrades of major infrastructure.

That's what GP100 is for.

 

[itsmydamnation]

yes and it says nothing, GCN could also be limited by the throughout of the ROP cache, you wouldn't know from that graph. But 290x vs 980 for example throughput of random texture is linear with width*clock so performance is consistent in that regard between maxwell and hawaii . colour compression only helps with a portion of data in a frame, there is still lots of shading and geometry data. factor this in with the difference in number of shaders and hawaii has a bandwidth advantage over a 980 per shader.

Now putting that aside it still doesn't say that 290 is bandwidth bottle necked it still has far more random texture throughput then a 980 with the same number of rops, which means 980 must be horribile castrated for 290 to be bottleneck... but it isn't..

If 290x is so bandwidth limited why are we seeing such improvements with DX12 and Async, because it isn't. Its DX11 driver limited, frontend limited, tesselation limited (not always) and on the high end requires a very large amount of concurrency in the sharders to get high utilization. DX12 removes two of those issues, if it was bandwidth bottle necked DX12 wouldn't be helping.

 

[JDG1980]

Why people believe GP104 will be 300mm2 ?? do we have any leak for that size ?? Because if Polaris 10 will be 230mm2, im expecting GM104 at 250mm2 range.

GF104: 332 mm^2
GK104: 294 mm^2
GM204: 398 mm^2

I expect GP104 to fit somewhere into this range. If it was substantially smaller than 300 mm^2, that would be unprecedented.

Polaris 10 is a Pitcairn-sized chip; its Nvidia counterparts would be the 6-series, not the 4-series.