AMD Fiji Officially Revealed at Computex; First Clear Pictures of GPU Core Emerge and New Teaser for Upcoming GPU

[Opcode]

There's a lot of measurement pics circulating. here's another one:

 

*edit* post above mine is the one I remember, which is why I said earlier its smaller than GM200.

 

Hard to properly measure a chip covered in goop. Although notice how his length of the die is about the same while facing the camera meanwhile the width of the die is better calculated by the image above where it too is facing the camera. I would say the above image is much more accurate since the length of the die is pretty much the same in both estimates. It's hard to get a width calculation based on the image you've posted due to the chip laying on an angle away from the camera. I would place my guess at 575-600 mm². I just don't see 480 mm² being a possibility that WCCFtech estimates. We could be looking at a chip the same exact size (give or take) as GM200.

[Decon]

has a chip similar to this "dual tonga" idea ever ben shipped?

[Opcode]

has a chip similar to this "dual tonga" idea ever ben shipped?

What do you mean by "dual Tonga"? A single big massive core? If so then yes the TITAN is another example big die GPU. On AMD stance Hawaii was a fairly big core but it was also conservative and nothing as big as Fiji. 100-200 mm² might not seem like a lot but under the hood that's millions of extra transistors.

[Decon]

What do you mean by "dual Tonga"? A single big massive core? If so then yes the TITAN is another example big die GPU. On AMD stance Hawaii was a fairly big core but it was also conservative and nothing as big as Fiji. 100-200 mm² might not seem like a lot but under the hood that's millions of extra transistors.

from the previous comments I got the impression that "Dual Tonga" meant AMD had somehow fused two Tonga dies to make Fiji.  As for big GPU cores I think the GTX 295 chips still reign supreme:

 

[osbios]

Another thing to take into consideration is the forthcoming of DirectX 12 which introduces support for hardware tiling. Which means entire textures no longer need to be loaded into virtual memory except just bits and pieces of textures (tiles) stored in page tables as needed. In reality by the time 4GB becomes an actual issue next generation cards will already be on the market. These cards aren't sold as future proof products but as high end enthusiast grade products for those looking for more for their dollar.

 

That's called sparse memory. (also calling it tilled just ask for misunderstandings -> "tilled rendering")

Rage was already using sparse memory via OpenGL. And that was released end of 2011.

Looked it up and apparently DirectX 11.2 already supports sparse memory. https://developer.nvidia.com/content/taking-advantage-directx112-tiled-resources

 

Considering the die size, isn't fury also manufactured using a smaller node? Or is it sill on 28nm?

[Decon]

That's called sparse memory. (also calling it tilled just ask for misunderstandings -> "tilled rendering")

Rage was already using sparse memory via OpenGL. And that was released end of 2011.

Looked it up and apparently DirectX 11.2 already supports sparse memory. https://developer.nvidia.com/content/taking-advantage-directx112-tiled-resources

 

Considering the die size, isn't fury also manufactured using a smaller node? Or is it sill on 28nm?

I believe it's still 28nm; same as Maxwell (and Kepler)

[Opcode]

from the previous comments I got the impression that "Dual Tonga" meant AMD had somehow fused two Tonga dies to make Fiji.  As for big GPU cores I think the GTX 295 chips still reign supreme:

 

geforce-gtx-295-pcb-detail-02.jpg

You could look at it from that point of view although really it's Tonga doubled (scaled up).

 

That's called sparse memory. (also calling it tilled just ask for misunderstandings -> "tilled rendering")

Rage was already using sparse memory via OpenGL. And that was released end of 2011.

Looked it up and apparently DirectX 11.2 already supports sparse memory. https://developer.nvidia.com/content/taking-advantage-directx112-tiled-resources

 

Considering the die size, isn't fury also manufactured using a smaller node? Or is it sill on 28nm?

Should still be on 28nm as everyone (except Samsung which doesn't have theirs commercially available) ditched 20nm to work on 14nm.

[Decon]

You could look at it from that point of view although really it's Tonga doubled (scaled up).

 

 

 

I fail to understand what you mean by "tonga doubled (scaled up)"

[Decon]

For those interested in predicting die size from pictures:

 

[Ex14]

I for one am interested in just seeing the cards performance. AMD has been teasing this for way too long what's worse is the release of the 980ti. I honestly hope it performs competitively for whatever price they're gonna be charging for their new cards. Also am hoping those rebadge rumors are fake. I'm not sure how competitive they can be with such old cards...

[Decon]

I for one am interested in just seeing the cards performance. AMD has been teasing this for way too long what's worse is the release of the 980ti. I honestly hope it performs competitively for whatever price they're gonna be charging for their new cards. Also am hoping those rebadge rumors are fake. I'm not sure how competitive they can be with such old cards...

everything except the new Fury card(s) are basically confirmed to be rebrands with increased clocks/memory

[Briggsy]

so what are the dimensions being measured? the logic die or the ram die? I ask because the logic die is slightly bigger than the Memory dies in the picture below. AMD says 5mmx7mm, while hynix says 5.48x7.29mm. Is hynix talking about the memory or the logic die? 

 

[Decon]

so what are the dimensions being measured? the logic die or the ram die? I ask because the logic die is slightly bigger than the Memory dies in the picture below. AMD says 5mmx7mm, while hynix says 5.48x7.29mm. Is hynix talking about the memory or the logic die? 

 

 

probably logic die

[Ex14]

everything except the new Fury card(s) are basically confirmed to be rebrands with increased clocks/memory

Still gonna be crossing my fingers. I have a 280X and its barely keeping pace with the 970s not sure how that will pan out if and is gonna pit it against the 970s. Unless they did some voodoo majicks and now its trading punches with it.

[Opcode]

I fail to understand what you mean by "tonga doubled (scaled up)"

Means they basically took Tonga and scaled up (doubling) it hardware wise. A full fat Tonga core has 2048 shaders meanwhile Fiji has 4096 etc.

 

so what are the dimensions being measured? the logic die or the ram die? I ask because the logic die is slightly bigger than the Memory dies in the picture below. AMD says 5mmx7mm, while hynix says 5.48x7.29mm. Is hynix talking about the memory or the logic die? 

 

AMD could be measuring just the stacked DRAM as they are comparing it to a GDDR5 module. Plus it would be a little weird to list decimal places on their slides.

 

 

If that is the case then it would be best to just measure both ways to get a range of likely die sizes (one at 5x7 and one at 5.48x7.29mm).

[Briggsy]

For those interested in predicting die size from pictures:

 

I copied and pasted a bunch of color modified HBM chips over the gpu die (I know, really ghetto methodology). using rough math and assuming the dies are 5.48mmx7.29mm, the gpu is definitely bigger than 600mm^2 because 12 full HBM's is already 480mm^2 at that size . If going by 5mmx7mm, I get roughly 540-550mm^2.

 

Also, ignore the grid. its not bang on.

[Decon]

I copied and pasted a bunch of color modified HBM chips over the gpu die (I know, really ghetto methodology). using rough math and assuming the dies are 5.48mmx7.29mm, the gpu is definitely bigger than 600mm^2 because 12 full HBM's is already 480mm^2 at that size . If going by 5mmx7mm, I get roughly 540-550mm^2.

 

Also, ignore the grid. its not bang on.

 

Hmm interesting...

[Guest Strangerbob]

anyone else already sick and tired of this card?

 

and realized they aint buying a 800$ card anyway?

[Gdourado]

anyone else already sick and tired of this card?

and realized they aint buying a 800$ card anyway?

Can't be 800 with the 980ti costing 650...

It would be commercial suicide for amd

[Guest Strangerbob]

Can't be 800 with the 980ti costing 650...

It would be commercial suicide for amd

 

Not if its at least 2% faster then the titan X across the board...   

 

The fastest card can cost whatever it wants and it will sell. Nvidia proved this.

 

 

Still, i get your point. And i wouldnt buy it even if its 550$. Never paying more than 400$ for a video card - those greedy cunts can go fuck themselves sideways.

[Humbug]

Can't be 800 with the 980ti costing 650...

It would be commercial suicide for amd

All our discussions about pricing are meaningless until we see the performance. $800 is ok if it has the performance for it, if it doesn't I doubt AMD will price it that high. The truth is we do not know about performance yet and too many people here have read too many rumors and think they know about this GPU.

[Linusinium]

Luke is too nice, Linus would have done the old "Sapphire told us not to enter this room but you know what, I'm just going to walk in there and look around" routine and gotten himself an exclusive scoop:

 

[Doobeedoo]

I want that on my key chain, or my belt, or necklace!

[Linusinium]

 

"That means that we can create a small form factor graphics card that is the fastest in the world."

4:52 of the presentation video. Not much room to back from those words. I'm boarding the 4K train soon with a ~$1500 GPU budget, now it's up to you AMD.

[Decon]

Dual Fury X please