[WCCF] AMD Radeon R9 390X Pictured

[Silverforce11]

I did my best... *shrug*

This is the size relative to HBM, HBM = shiny. Interposer is the dark base all around the entire GPU + HBM.

It's a monster.

Suddenly statements like "a lot faster than leaks indicate" from journalists with first hand testing makes sense.

 

[Silverforce11]

I agree.

Shinies!

Come on, it's obvious.

 

[MrTeal]

Shinies!

Come on, it's obvious.

The shininess will depend on what side of the chip you're looking at. The backside will be much less iridescent and more uniform than the bottom.

 

[96Firebird]

How big are the pins usually? I'll just measure off those, even though they are pretty blurry...

The problem with the shinies being the HBM is that I can't get it to fit the rectangular shape of HBM, but I'll keep trying...

 

[MrTeal]

How big are the pins usually? I'll just measure off those, even though they are pretty blurry...

What pins?

 

[Silverforce11]

Anyone care to speculate on the performance now, knowing how freaken massive it is?

I'm gonna speculate right here.

Hawaii = 438mm2
GK110 = 561mm2
GM204 = 398mm2
GM200 = 601mm2

Fiji = ~800mm2 (900mm2 is right at the max production capability IIRC)

Hawaii is ~10% behind GM204, it's also ~10% larger. Thus, GCN1.1 has a 20% Perf/mm2 deficit compared to GM204.
Titan X is ~48% faster than Hawaii, it's also 37% larger. Thus, GCN1.1 has a similar Perf/mm2 deficit compared to GM204.

Close enough against GM204 & GM200, let's say GCN1.1 needs 20% larger die area to match Maxwell 2 performance.

601mm2 x 1.2 = 720mm2.

Assuming zero uarch changes (which isn't true, but lets assume no improvements in perf/mm2), Fiji should already be faster than Titan X.

Produced on GloFo (confirmed, GF has publicly mentioned it), as seen with Carrizo, ~20-30% efficiency gains compared to TSMC 28nm.

Add in uarch changes & HBM advantage on top.. beast mode!

 

[MrTeal]

Anyone care to speculate on the performance now, knowing how freaken massive it is?

I'm gonna speculate right here.

Hawaii = 438mm2
GK110 = 561mm2
GM204 = 398mm2
GM200 = 601mm2

Fiji = ~800mm2 (900mm2 is right at the max production capability IIRC)

Hawaii is ~10% behind GM204, it's also ~10% larger. Thus, GCN1.1 has a 20% Perf/mm2 deficit compared to GM204.
Titan X is ~48% faster than Hawaii, it's also 37% larger. Thus, GCN1.1 has a similar Perf/mm2 deficit compared to GM204.

Close enough against GM204 & GM200, let's say GCN1.1 needs 20% larger die area to match Maxwell 2 performance.

601mm2 x 1.2 = 720mm2.

Assuming zero uarch changes (which isn't true, but lets assume no improvements in perf/mm2), Fiji should already be faster than Titan X.

Produced on GloFo (confirmed, GF has publicly mentioned it), as seen with Carrizo, ~20-30% efficiency gains compared to TSMC 28nm.

Add in uarch changes & HBM advantage on top.. beast mode!

Knowing this how? No offense to Firebird, but his initial measurement based on a blurry photo of the incorrect part of the chip isn't exactly proof positive.

 

[iiiankiii]

You can see the core is raised up against the interposer at the top. I think those lighter/brown areas are the HBM

That thing is massive.. no pun intended.

 

[96Firebird]

What pins?

The rows on top and bottom, they are tiny but would give me another reference point. I'm assuming they're pins.

Last attempt, tried my best to fit the HBM size to a shiny part.

 

[Silverforce11]

There's the other calculation also:
http://forums.anandtech.com/showpost.php?p=37445180&postcount=1751

>700mm2.

 

[HurleyBird]

Shiny part is interposer

 

[Silverforce11]

The rows on top and bottom, they are tiny but would give me another reference point. I'm assuming they're pins.

Last attempt, tried my best to fit the HBM size to a shiny part.

You got HBM over-measured on length, as you can see from .vodka's die shot, the HBM chip does not extend (small offset) all the way to the end/align with the GPU die, if you know what I mean.

Here, the HBM rectangle shape lines up very clear.

 

[MrTeal]

The rows on top and bottom, they are tiny but would give me another reference point. I'm assuming they're pins.

Last attempt, tried my best to fit the HBM size to a shiny part.

Ah. Those aren't pins, they're ceramic capacitors.

Look at the black rectangles, the top ones are slightly offset from the edge of the GPU die, just like in the shot with all the thermal paste.

 

[96Firebird]

First off, I see I'm missing some images that others are posting... For some reason, Opera doesn't show images from a lot of hosting sites like tinypic or photobucket, I think because the image opens up in a webpage...?

I'll see what I can do with the blue background pic. And I now think the dark part is the HBM.

/me fires up SolidWorks again...

 

[Silverforce11]

Shiny part is interposer

The interposer looks dull compared to the metallic HBM surface.

What a silly non-conclusive leak from Lisa! More speculation until the 16th...

 

[RaulF]

Shiny part is interposer

After looking at DDH picture, i think you might be right. The black squares seem to pop more.

Found this image that might agree that the interposer is the shiny part.

 

[MrTeal]

The interposer looks dull compared to the metallic HBM surface.

The die looks shiny, but it's the same uniform shininess you see with the back side of a flip chip when it's been mounted. That iridescence is what you see on the top side of the chip. The GPU and HBM on the chip have the top side down (flip chip).
Compare the top side on this Intel Westmere chip

With the one of it mounted flip chip onto the substrate. The backside just looks like a shiny mirror.

 

[flopper]

Anyone care to speculate on the performance now, knowing how freaken massive it is?

Fury X faster than TX
Fury Pro equal or slightly lower to 980ti
If thats accurate they can price the furu pro between 980 and 980ti and slaughter both.

 

[Silverforce11]

Interposer size is not similar if you look at the presentation image and the TIM covered die shot. The positioning of HBM means that the interposer at one end is longer than the interposer gap between HBM modules.

Looking at the die Lisa is holding, the 4 shiny seem similar in size. The black area match the size of the interposer, the gap between 2 shiny is smaller than the other one.

 

[96Firebird]

Last attempt for real this time, I swear.

Like .vodka said in the other thread, it is really easy to be just a half a mm off and it'll change everything drastically. Also, the method I am using works best when the object is completely perpendicular to the camera, which of course these shots are not.

Oh well, I gave it my best. Do what you want with my measurements...

The only conclusion I am drawing from this is that it is over 500mm².

 

[eRacer]

Anyone care to speculate on the performance now, knowing how freaken massive it is?

Fiji = ~800mm2

It won't be near 800 mm^2. That picture didn't measure the core if it has the heatspreader on it. I believe a heatspreader over the GPU core is required to reach the same height at HBM so there isn't an air gap between the CPU and cooler because the HBM is taller than the core. Can't really tell from the pic if that is the GPU core or heatspreader. So either the width and height measurements are pretty far off or the GPU core doesn't extend all the way out to the edges of the heatspreader.

 

[Azix]

HBM is the dark. The shiny are squares, not rectangles

 

[iiiankiii]

HBM is the dark. The shiny are squares, not rectangles

I don't know. The dress looks blue and black to me. Definitely not gold and white. oh wait... what are we talking about?

Kidding aside, I agree. HBM seems to be the dark ones.

 

[Silverforce11]

Last attempt for real this time, I swear.

Like .vodka said in the other thread, it is really easy to be just a half a mm off and it'll change everything drastically. Also, the method I am using works best when the object is completely perpendicular to the camera, which of course these shots are not.

Oh well, I gave it my best. Do what you want with my measurements...

The only conclusion I am drawing from this is that it is over 500mm².

With that contrast of the image, yes you guys are right. It looks like shiny = interposer and dark = HBM.

 

[HurleyBird]

It won't be near 800 mm^2. That picture didn't measure the core. It measured the heatspreader over the CPU core which is designed to reach the same height at HBM so there isn't an air gap between the CPU and cooler because the HBM is taller than the core.

Certainly doesn't look like a heatspreader to me.