[WCCF] AMD Radeon R9 390X Pictured

[.vodka]

Oh, excellent. I'd missed that bit of info. With a more precise known measurement we can do better:



7.29mm is the base, known measurement. This time I also applied a sharpening filter that reduces lens blur, makes it easier to discern where the edges are.

GPU: 25.72*27.17mm = 698.81mm²
Interposer: 41.21*28.72mm = 1183.55mm²


If we round that:
GPU: 26*27mm = 702mm²
Interposer: 41*29mm = 1189mm²


Give or take a mm, margin of error we're dealing with here:

GPU: 25*26mm = 650mm²
Interposer: 40*28mm = 1120mm²


But yeah, this is a big, big GPU either way. Should be an interesting launch.

we're talking mm here so might be right. Errors on where the chips end and start included, could be mm difference one way or the other. Wish it was top down cleaned shot.

This person must have been in some kind of rush to take a pic like that, or they are teasing and this is not fiji. Or they don't want to reveal the name of what we call fiji. smh

soon....

Yeah, I also wish we had a better picture to speculate on. But it is what it is. At least this quite handy tool PS has lets us take perspective into account.

It isn't much longer now to see what that thing is capable of.

 

[twjr]

Interposer costs = $1 / 100mm2

http://electroiq.com/blog/2012/12/lifting-the-veil-on-silicon-interposer-pricing/

Sure its cheap. But adding chips on the empty sides is going to add ~50% to the area of the interposer. Based on the guesses above that will add ~$5 to the cost of the interposer.

If it were possible to mount the HBM chips short-side on the GPU then they could fit 8 chips by only lengthening the interposer by ~5mm for a lot less money.

Saving ~$5/package is big money in a mass produced thing. Look at Intel using TIM instead of solder.

If anything this probably further confirms there are only 4GB models unless AMD has a totally different package with a different layout for the 8GB model.

Edit: mind you until we have reviews we have no idea how 4GB of HBM performs. TBH at the moment all we know is this is an F****** big chip.

 

[Tuna-Fish]

What if they could add six layers in each HBM stack?

Each HBM stack has exactly 8 channels, regardless of height. The first-gen HBM has two channels per layer, later ones will have one channel per layer. Each stack has 4 or 8 memory layers.

 

[Azix]

the revised are way too high.

what is the mold and is it visible?

 

[Silverforce11]

Anything around 600mm2 would be freaken epic already.

Bigger than that and they are not playing around but going for the performance crown and more.

 

[maddie]

Each HBM stack has exactly 8 channels, regardless of height. The first-gen HBM has two channels per layer, later ones will have one channel per layer. Each stack has 4 or 8 memory layers.

Where did you see this? HBM1?

 

[Silverforce11]

Where did you see this? HBM1?

AMD/Hynix said there's no 4 stack limit, it can go higher, but they did not specify what's going to be on Fiji, when asked about the 4GB vram limit, the response was "we have something for that".

 

[maddie]

AMD/Hynix said there's no 4 stack limit, it can go higher, but they did not specify what's going to be on Fiji, when asked about the 4GB vram limit, the response was "we have something for that".

I'm asking about the # of layers not # of stacks. I never read anywhere that 8 layers is possible on HBM-1.

 

[chimaxi83]

Giant chip from AMD after all this time? Yes please! I hope the performance is there.

 

[DiogoDX]

Where did you see this? HBM1?

http://videocardz.com/55259/sk-hynix-shows-off-hbm1-and-hbm2

 

[3DVagabond]

AMD/Hynix said there's no 4 stack limit, it can go higher, but they did not specify what's going to be on Fiji, when asked about the 4GB vram limit, the response was "we have something for that".

I think they said the number of stacks wasn't limited to 4, not that you could have more than 4 per stack.

Edit: Didn't see the post above this one.

 

[Silverforce11]

I think they said the number of stacks wasn't limited to 4, not that you could have more than 4 per stack.

4 or 8 layers was presented by Hynix.

As above.

So 4/8GB is a possibility on 4 stacks.

 

[gamervivek]

If it is closer to 700mm2 than 600, the $1200 for Fiji samples landing in India would start making a whole lot more sense.

 

[Tuna-Fish]

I'm asking about the # of layers not # of stacks. I never read anywhere that 8 layers is possible on HBM-1.

8 layers is theoretically possible, and was originally on Hynix roadmap, but as I understand they are now not planning to make it until HBM2.

To the size speculation, imho it's hard to see the size of the die itself from that pic.

One of the only actual pieces of information, as opposed to random slides from unverified origins that we have about future AMD products is this slide from synapse design:

That tells us that they have made design work on a >500mm^2 GPU for AMD. I expect that Fiji is >500mm^2, <550mm^2, and made at TSMC.

 

[3DVagabond]

4 or 8 layers was presented by Hynix.

As above.

So 4/8GB is a possibility on 4 stacks.

Did you not see my edit?

 

[gamervivek]

It's possible that the >500mm2 GPU is Hawaii replacement and second in line after Fiji. Of course it's also possible that these calculations are overreaching.

 

[LTC8K6]

It looks like the memory and the gpu need to be the same height in that setup.

It also looks like someone could simply have mocked something up to look like a GPU with HBM stacks.

 

[.vodka]

Yes. The angle of the picture here doesn't help, a little variation of 0.29mm on the reference measurement resulted in 100mm² more on the GPU measurement. One single mm as margin of error on both dimensions is ±25-50mm² on the final die size.

What doesn't help either is all the thermal paste obscuring the die edges, there's a lot of guessing that introduces even more error...

So take these with caution, but we can all agree this is a *huge* GPU by AMD standards, easily over 500 or 550mm² as the minimum. This is nvidia die size territory and it's a first for AMD. The synapse slide at least confirmed that months ago and we now have a picture of what could be that >500mm² chip.

 

[Elfear]

Didn't some leak or AMD rep say that GPUs with HBM would require an IHS to make sure everything was on the same plane? If that is the case could the die size look overstated in that picture because it's actually the IHS?

 

[destrekor]

Didn't some leak or AMD rep say that GPUs with HBM would require an IHS to make sure everything was on the same plane? If that is the case could the die size look overstated in that picture because it's actually the IHS?

That's a very good point! It is quite logical that the HBM and GPU will not have perfectly level die heights.

 

[Vesku]

Didn't some leak or AMD rep say that GPUs with HBM would require an IHS to make sure everything was on the same plane? If that is the case could the die size look overstated in that picture because it's actually the IHS?

Then we wouldn't be seeing the HBM would we? This kind of goopy thermal paste job would make a lot of sense if it's under an IHS they removed.

 

[Elfear]

Then we wouldn't be seeing the HBM would we? This kind of goopy thermal paste job would make a lot of sense if it's under an IHS they removed.

Good question. I would assume the IHS would cover the HBM and GPU but the thermal paste looks a lot goopier than what you'd see under the lid of a 4770k. The paste looks more like what you'd find on top of the IHS under a heatsink.

 

[Vesku]

Good question. I would assume the IHS would cover the HBM and GPU but the thermal paste looks a lot goopier than what you'd see under the lid of a 4770k. The paste looks more like what you'd find on top of the IHS under a heatsink.

4 other small chips to cover, though, and choice of paste. The factory paste jobs on my GPUs when I've disassembled them, direct die to cooler, have been more like the Intel method. Looks like AMD decided not to pinch pennies on thermal materials for this GPU.

 

[RussianSensation]

PCGamesHardware:
http://www.pcgameshardware.de/AMD-R...750-Euro-Preisschild-1160525/galerie/2381250/

1150mhz GPU clock
8GB HBM1
640GB/sec HBM1

May 27th BIOS

1150mhz clocks with 4096 shaders would make the card up to 65%+ faster than an R9 290X at 4K. Probably fake since sounds way too good to be true.

EDIT: Oh looks like I was beaten 4 pages ago. Nice LTC8K6!

leak is right on time...

https://lh4.googleusercontent.com/-...AAALJs/62rqQEaLQ-g/w2178-h1225-no/desktop.jpg

 

[ocre]

already discussed that

http://forums.anandtech.com/showthread.php?t=2433358

It was concluded to be fake

edit: as for your edit
we had a whole thread that dealt with it. that is the link i posted