[headd]

Raja says they
Quote:

Originally Posted by prjindigo

They said, at the conference, that it's HBM... that they're not bothering with trying to make two different memory architectures.

Where is this 256bit GDDR5 8GB IDIOCY coming from?!? 8GB can only be done at 512bit!

Really?

Raja said polaris wont USE HBM2 because its not ready for mainstream
https://youtu.be/4qJj1ViyyPY?t=594

 

[iLeakStuff]

Quote:

Originally Posted by prjindigo

... Specifications: Except at the press conference He said they weren't making any GDDR5 Polaris.

Incorrect. Raja said that they will use GDDR5 where "it makes financially sense". To me it means price/performance cards.

 

[prjindigo]

Quote:

Originally Posted by headd

Raja says they
Really?

Raja said polaris wont USE HBM2 because its not ready for mainstream
https://youtu.be/4qJj1ViyyPY?t=594

Is today weed day or something? HBM is not HBM2. Never said HBM2.

 

[prjindigo]

Quote:

Originally Posted by iLeakStuff

Incorrect. Raja said that they will use GDDR5 where "it makes financially sense". To me it means price/performance cards.

Except that he also said it wasn't going to be used on Polaris.

I've inspected the "sources" listed by original poster and found nothing referring to the 256bits or 8gb or GDDR5.

Which makes this thread a troll thread.

 

[headd]

Quote:

Originally Posted by prjindigo

Is today weed day or something? HBM is not HBM2. Never said HBM2.

Maybe you have too much of it.

 

[lolfail9001]

Quote:

Originally Posted by prjindigo

Except that he also said it wasn't going to be used on Polaris.

I've inspected the "sources" listed by original poster and found nothing referring to the 256bits or 8gb or GDDR5.

Which makes this thread a troll thread.

http://semiaccurate.com/forums/showpost.php?p=258453&postcount=513

Here, your source for 256 [email protected] of GDDR5 on Polaris 10.

Now, amount of RAM is only sourced by those benchmarks and they look exploitable, but it sounds reasonable enough.

To rest: come on, don't kid yourselves, we'll be very lucky if reference 480x will be better than 390x in real world performance, rather than performance per watt or dollar.

 

[iLeakStuff]

Quote:

Originally Posted by lolfail9001

http://semiaccurate.com/forums/showpost.php?p=258453&postcount=513

Here, your source for 256 [email protected] of GDDR5 on Polaris 10.

Now, amount of RAM is only sourced by those benchmarks and they look exploitable, but it sounds reasonable enough.

To rest: come on, don't kid yourselves, we'll be very lucky if reference 480x will be better than 390x in real world performance, rather than performance per watt or dollar.

Nobody said this was 480X. It might be, but it could be 490X as well.
280X was not targeted towards GTX 680, and you bet AMD have a Polaris 10 waiting to be released to counter GTX 1080.

 

[tajoh111]

Quote:

Originally Posted by iLeakStuff

Why do I have to repeat myself:

A) Its only +30% core performance over GCN required.
B) AMD showed a presentation where Polaris 10 beat Fury X. The one in OP with 2304 cores is Polaris 10. One would assume its close to Fury X too.

30% is tremendous, particularly given the die size of the chip(232mm2). In addition because of the performance of the fury x and gtx 980 ti, bandwidth becomes an issues after a certain performance level.

The gtx 980 performance over the gtx 780 ti was possible because of the increase in clock speed and hardware based global illumination hardware put in the gtx 980. In addition it kicked out the double precision performance and used a much smaller memory controller than the gtx 780 ti.

As far as pure terraflops go, the gtx 980 is only 10% lower than a gtx 780 ti. So add 15% efficiency gains + games where global illumination is present in games and you get the difference between the gtx 980 and 780 ti.

If we look at the increase in transistor number between the 6970 and 7870, the 7870 outperformed the 6970 by 10% but the increase in transistors was near proportional to this. The 7870 had 2.8 billion in transistors, and the 6970 had 2.64 billion transistors. In addition, the double precision of pitcairns was nerfed further compared to the 6970. Compared to the 440mm2 die of hawaii compared to the the 232mm2 die size of polaris 10, I wouldn't be surprised if they performed similarly. Add in the 390 variants are clocked rather high for a GCN based architecture, it wouldn't surprise me if they performed similarly. The hitman demo did not show all of it's settings and many people said that a 390x could achieve those results.

Achieving fury X or faster results would mean AMD would have to increase their per transistor performance by 50%. This just doesn't seem possible because Fury X is running a CLC, and it has already sacrifaced its FP64 and polaris 10 is stilling using a GCN derivative.

 

[lolfail9001]

Quote:

Originally Posted by iLeakStuff

Nobody said this was 480X. It might be, but it could be 490X as well.
280X was not targeted towards GTX 680, and you bet AMD have a Polaris 10 waiting to be released to counter GTX 1080.

It has to be 480X, because Vega won't be R9 5xx and almost certainly won't be another named chip above the rest (Fury was an exception, not the rule, that is not hard to figure out), so it just leaves 490/490x for Vega 11/10 respectively. Also, 7970 and 680 were in general trading blows and had same MSRP, so i would say 7970 was aimed straight at 680.

Now, the problem with Polaris 10 is that we actually have 0 clue how well it performs, because nobody has ever revealed settings used in this demo run of Hitman at Polaris 10 tease.

What we know is that it won't perform better than margin of error over Fury X.

What we see in practice here is that Polaris 10 (or at least cut down version of it) has less shading units than 290/390, so full version almost certainly has less shading units than 290x/390x as well.
Can AMD squeeze out 30% gain out of thin air without complete architecture overhaul? Doubt it. Because there is no complete architecture overhaul.

 

[iLeakStuff]

Quote:

Originally Posted by lolfail9001

It has to be 480X, because Vega won't be R9 5xx and almost certainly won't be another named chip above the rest (Fury was an exception, not the rule, that is not hard to figure out), so it just leaves 490/490x for Vega 11/10 respectively. Also, 7970 and 680 were in general trading blows and had same MSRP, so i would say 7970 was aimed straight at 680.

Now, the problem with Polaris 10 is that we actually have 0 clue how well it performs, because nobody has ever revealed settings used in this demo run of Hitman at Polaris 10 tease.

What we know is that it won't perform better than margin of error over Fury X.

What we see in practice here is that Polaris 10 (or at least cut down version of it) has less shaders than 290/390, so full version almost certainly has less shaders than 290x/390x as well.
Can AMD squeeze out 30% gain out of thin air without complete architecture re-haul? Doubt it. Because there is no complete architecture re-haul.

Vega will have HBM2, so AMD could do another round with fancy names like Fury X to make it special. Vega wont be out until 2017 since thats the GPU to take on Titan P and GTX 1080Ti. So AMD need something to counter GP104 from June and til 2017, and since AMD said Polaris 10 and Polaris 11 are the only chips out this year, with Polaris 11 being low end, Polaris 10 are our only option.

A possible scenario here for you sceptics is Polaris 10 GPUs ranging all the way from 2000 shaders up to 3000 shaders, to be able to make GPUs from 480 up to 490X. Thats a possibility.
Or Polaris 11 extends up to 480X, while 490 and 490X is Polaris 10.

I too am a little sceptical about this performance increase per core, but it is certainly possible. There are a lot of changes coming with Polaris. We shall see.

 

[headd]

Quote:

Originally Posted by iLeakStuff

Vega wont be out until 2017 since thats the GPU to take on Titan P and GTX 1080Ti. So AMD need something to counter GP104 from June and til 2017, and since AMD said Polaris 10 and Polaris 11 are the only chips out this year, with Polaris 11 being low end, Polaris 10 are our only option.
.

They will not counter GP 104.
Polaris 10 and 11 are low-end and mainstream of 14/16nm era.Gp 104 is another Tier.Vega will counter Gp 104.
AMD will have better cards in low-end and up to GTX980 performance -100-300USD market.
NV with Gp 104 will have 400-550USD market.

NV will not release Mainstream pascal untill end of year.

Polaris 10 and 11 will be better than anything NV curretly have up to GTX980.But nv will have GP 104 with no competition in 400-550USD market.

 

[Forceman]

Vega will counter GP100, and Polaris will counter GP104. There is no way AMD is going to concede the GP104 market to Nvidia until Vega comes out in 2017.

 

[headd]

How can 232mm2 2560SP polaris counter GP 104 with 20% or more performance Than 980TI?
Btw they will have entire Market Up TO GTX980 with far better 14nm cards vs old 28nm NV cards.
I am sure its better for AMD leave 2 cards in 400-550USD market to NV and have better cards in rest market.

 

[Pyrotagonist]

Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106
Polaris 11 (Cape Verde replacement) ~ GP108

That, I think, is what AMD is going for.

 

[headd]

Quote:

Originally Posted by Pyrotagonist

Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106
Polaris 11 (Cape Verde replacement) ~ GP108

That, I think, is what AMD is going for.

YEP

 

[Forceman]

Quote:

Originally Posted by headd

How can 232mm2 2560SP polaris counter GP 104 with 20% or more performance Than 980TI?
Btw they will have entire Market Up TO GTX980 with far better 14nm cards vs old 28nm NV cards.
I am sure its better for AMD leave 2 cards in 400-550USD market to NV and have better cards in rest market.

Couple of thoughts:

They already demonstrated a Polaris card that they say was showing Fury X class performance (1440p Ultra Hitman).
Do we know it is only 232 mm^2?
GP104 probably won't get to 20% more than GTX 980 Ti (more likely 10-15% max).
Why do you think Nvidia won't have 16nm cards below GP104?
Quote:

Originally Posted by Pyrotagonist

Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106
Polaris 11 (Cape Verde replacement) ~ GP108

That, I think, is what AMD is going for.

Where's the Fiji replacement?

 

[Serandur]

Quote:

Originally Posted by prjindigo

Where is this 256bit GDDR5 8GB IDIOCY coming from?!? 8GB can only be done at 512bit!

That is so untrue. 8 Gb (1 GB) GDDR5 chips have been an option for quite some time. A 256-bit GDDR5-equipped memory bus handles 8 chips in standard mode and can handle 16 chips in clamshell mode. Eight 8 Gb (1 GB) chips = 8 GBs and if they go for clamshell mode, they could even fit 16 GBs on there if they wanted to.

 

[Pyrotagonist]

Quote:

Originally Posted by Forceman

Couple of thoughts:

They already demonstrated a Polaris card that they say was showing Fury X class performance (1440p Ultra Hitman).
Do we know it is only 232 mm^2?
GP104 probably won't get to 20% more than GTX 980 Ti (more likely 10-15% max).
Why do you think Nvidia won't have 16nm cards below GP104?
Where's the Fiji replacement?

Probably won't be one until Navi. Fiji is an oddball anyway, doesn't easily fit into the hierarchy.

 

[gamervivek]

Quote:

Originally Posted by Pyrotagonist

Vega 10 (Hawaii replacement) ~ GP100
Vega 11 (Tahiti replacement) ~ GP104
Polaris 10 (Pitcairn replacement) ~ GP106
Polaris 11 (Cape Verde replacement) ~ GP108

That, I think, is what AMD is going for.

+1 on the AMD side, though holding out hope for a Fiji sized chip.

The only hitch I see in predicting the new cards performance is the price they are shipping to India on zauba. Baffin XT, the cape verde replacement's best configuration, is 60% of Fiji XT right now which makes the bigger chip end up in the 500$ and above zone which is too much for something only as good as Hawaii/Grenada.i

 

[tajoh111]

Quote:

Originally Posted by Forceman

Couple of thoughts:

They already demonstrated a Polaris card that they say was showing Fury X class performance (1440p Ultra Hitman).
Do we know it is only 232 mm^2?
GP104 probably won't get to 20% more than GTX 980 Ti (more likely 10-15% max).
Why do you think Nvidia won't have 16nm cards below GP104?
Where's the Fiji replacement?

We don't have the settings for Hitman, just the resolution.

Releasing fury pro duo doesn't make sense if polaris is going to be as fast as fury x. Polaris 10 has been more or less confirmed to be a 232mm2 chip. Fury x has a CLC and is 600mm2 and no fp64. It going to be difficult to beat fury X compared to AMD earlier chips beating their flagship.

This years release schedule is likely

Polaris 10 to succeed pitcairns and replace hawaii/tonga
Fury X is going to stay the fastest AMD chip
Polaris 11, is going to be a volume/laptop part that replace cape verde.

AMD doesn't have the money Nvidia has so can't release as wide of a range of chips and they need to recover their BOM on fury x. To have FuryX and their initial investment on HBM be replaced so soon would be irresponsible with their limited resources, it something they can't afford to do. Taping out, designing and fabricating chips cost 100 million this generation.

a 2560 core card makes sense given tonga's specs.

Rough guesstimate, but take tonga's specs, times it by 1.333 and you should have what the specs of this card using ddr5. Tonga is 360mm2 so divide that by 2(to denote the shrink) and that's 180. Times this by 1.333 = 240mm2. Reduce specs accordingly, if per shaders performance is increase and thus cores become more complex and you start getting the specs Polaris 10 should have. This is only the cores those, texture units and ROPs will change between generations.